├── .gitignore
├── README.md
├── hand-estimator-using-caffemodel
├── README.md
├── handpose.py
├── model
│ └── pose_deploy.prototxt
└── test_out
│ ├── README.md
│ ├── hand_out.jpg
│ └── shake_out.jpg
├── pose-estimator-tensorflow
├── README.md
├── openpose_skeleton_sequence_drawer.py
├── openpose_tf.py
└── test_out
│ ├── README.md
│ ├── skeleton_sequence.jpg
│ └── skeleton_sequence_horizontal.jpg
├── pose-estimator-using-caffemodel
├── README.md
├── estimator_multi_people.py
├── estimator_single_person.py
├── model
│ ├── body_25
│ │ ├── keypoints_pose_25.png
│ │ └── pose_deploy.prototxt
│ ├── coco
│ │ ├── keypoints_pose_18.png
│ │ ├── pose_deploy_linevec.prototxt
│ │ └── pose_deploy_linevec_faster_4_stages.prototxt
│ └── mpi
│ │ ├── pose_deploy_linevec.prototxt
│ │ └── pose_deploy_linevec_faster_4_stages.prototxt
└── test_out
│ └── README.md
└── test_out
├── CSBDGS_out.jpg
├── Friends_out.jpg
├── Output-BODY_25.jpg
├── Output-COCO.jpg
├── Output-MPI.jpg
├── TBBT_out.jpg
├── WLWZ_out.jpg
└── webcam_out.gif
/.gitignore:
--------------------------------------------------------------------------------
1 | # Byte-compiled / optimized / DLL files
2 | __pycache__/
3 | *.py[cod]
4 | *$py.class
5 |
6 | # C extensions
7 | *.so
8 |
9 | # Distribution / packaging
10 | .Python
11 | build/
12 | develop-eggs/
13 | dist/
14 | downloads/
15 | eggs/
16 | .eggs/
17 | lib/
18 | lib64/
19 | parts/
20 | sdist/
21 | var/
22 | wheels/
23 | *.egg-info/
24 | .installed.cfg
25 | *.egg
26 | MANIFEST
27 |
28 | # PyInstaller
29 | # Usually these files are written by a python script from a template
30 | # before PyInstaller builds the exe, so as to inject date/other infos into it.
31 | *.manifest
32 | *.spec
33 |
34 | # Installer logs
35 | pip-log.txt
36 | pip-delete-this-directory.txt
37 |
38 | # Unit test / coverage reports
39 | htmlcov/
40 | .tox/
41 | .nox/
42 | .coverage
43 | .coverage.*
44 | .cache
45 | nosetests.xml
46 | coverage.xml
47 | *.cover
48 | .hypothesis/
49 | .pytest_cache/
50 |
51 | # Translations
52 | *.mo
53 | *.pot
54 |
55 | # Django stuff:
56 | *.log
57 | local_settings.py
58 | db.sqlite3
59 |
60 | # Flask stuff:
61 | instance/
62 | .webassets-cache
63 |
64 | # Scrapy stuff:
65 | .scrapy
66 |
67 | # Sphinx documentation
68 | docs/_build/
69 |
70 | # PyBuilder
71 | target/
72 |
73 | # Jupyter Notebook
74 | .ipynb_checkpoints
75 |
76 | # IPython
77 | profile_default/
78 | ipython_config.py
79 |
80 | # pyenv
81 | .python-version
82 |
83 | # celery beat schedule file
84 | celerybeat-schedule
85 |
86 | # SageMath parsed files
87 | *.sage.py
88 |
89 | # Environments
90 | .env
91 | .venv
92 | env/
93 | venv/
94 | ENV/
95 | env.bak/
96 | venv.bak/
97 |
98 | # Spyder project settings
99 | .spyderproject
100 | .spyproject
101 |
102 | # Rope project settings
103 | .ropeproject
104 |
105 | # mkdocs documentation
106 | /site
107 |
108 | # mypy
109 | .mypy_cache/
110 | .dmypy.json
111 | dmypy.json
112 |
113 | # Pyre type checker
114 | .pyre/
115 |
116 | test/1.jpg
117 | test/MrFang.mp4
118 | test/mySweet.jpg
119 | *.mp4
120 | *.caffemodel
121 | *.pb
122 | .idea
123 | __pycache__
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # OpenPose Rebuilt-Python
2 | Rebuilting the CMU-OpenPose pose estimatior using Python with OpenCV and Tensorflow.
3 | (The code comments are partly descibed in chinese)
4 |
5 | -------
6 | ## Pretrained-model Downloading
7 | In this work, I used both **caffemodel** and **tensorflow-graph-model**, you can download them [here](https://pan.baidu.com/s/1XT8pHtNP1FQs3BPHgD5f-A), Then place the pretrained models to corresponding directory respectively.
8 | #### *Examples:*
9 | - place `caffe_models\pose\body_25\pose_iter_584000.caffemodel` into `pose-estimator-using-caffemodel\model\body_25\`
10 | - place `caffe_models\hand\pose_iter_102000.caffemodel` into `hand-estimator-using-caffemodel\model\`
11 | - place `openpose graph model coco\graph_opt.pb` into `pose-estimator-tensorflow\graph_model_coco\`
12 |
13 | -------
14 | ## Requirements :
15 | - OpenCV > 3.4.1
16 | - TensorFlow > 1.2.0
17 | - imutils
18 |
19 | -------
20 | ## Usage:
21 | See the sub-README.md in sub-folder.
22 |
23 | -------
24 | ## BODY_25 vs. COCO vs. MPI
25 | - BODY_25 model is ***faster, more accurate***, and it includes foot keypoints.
26 | - COCO requires less memory on GPU (being able to fit into 2GB GPUs with the default settings) and it runs ***faster on CPU-only mode***.
27 | - MPI model is only meant for people requiring the MPI-keypoint structure. It is also slower than BODY_25 and far less accurate.
28 | #### *Output Format*
29 | **Body_25 in left, COCO in middle, MPI in right.**
30 |
31 |
32 | **See more Output Format details [here](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/output.md)**, and **Hand Output Format** included as well.
33 |
34 | -------
35 | ## Results Showing
36 | #### *Image test*
37 |
38 | 
39 |
40 |
41 | 
42 |
43 |
44 | 
45 |
46 |
47 | 
48 |
49 |
50 | #### *VideoStream test*
51 |
52 | 
53 |
54 |
55 | #### *Webcam-skeleton-drawer test*
56 | Script in `pose-estimator-tensorflow` folder.
57 |
58 | 
59 |
60 |
61 | 
62 |
63 |
--------------------------------------------------------------------------------
/hand-estimator-using-caffemodel/README.md:
--------------------------------------------------------------------------------
1 | ## Usage Examples :
2 | Put the test file (iamge or video) under the same directory
3 |
4 | - `python3 handpose.py --image=test.jpg`
5 | - `python3 handpose.py --video=test.mp4`
6 | - if no argument provided, it starts the webcam.
7 |
8 | ## Note
9 | - see more details [here](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/output.md#hand-output-format)
10 | - see the origin paper [here](https://arxiv.org/abs/1704.07809)
11 | - the code given can detect ***only one hand*** at a time.
12 |
13 | ## Result showing
14 |
15 | 
16 |
17 |
18 | 
19 |
20 |
--------------------------------------------------------------------------------
/hand-estimator-using-caffemodel/handpose.py:
--------------------------------------------------------------------------------
1 | import cv2 as cv
2 | import os
3 | import argparse
4 | import sys
5 | from imutils.video import FPS
6 | from pathlib import Path
7 |
8 |
9 | FilePath = Path.cwd()
10 | outFilePath = Path(FilePath / "test_out/")
11 | threshold = 0.2
12 | input_width, input_height = 368, 368
13 |
14 | proto_file = str(FilePath/'model/pose_deploy.prototxt')
15 | weights_file = str(FilePath/"model/pose_iter_102000.caffemodel")
16 | nPoints = 22
17 | POSE_PAIRS = [[0, 1], [1, 2], [2, 3], [3, 4], [0, 5], [5, 6], [6, 7], [7, 8], [0, 9],
18 | [9, 10], [10, 11], [11, 12], [0, 13], [13, 14], [14, 15], [15, 16], [0, 17],
19 | [17, 18], [18, 19], [19, 20]]
20 |
21 | # Usage example: python openpose.py --video=test.mp4
22 | parser = argparse.ArgumentParser(description='OpenPose for hand skeleton in python')
23 | parser.add_argument('--image', help='Path to image file.')
24 | parser.add_argument('--video', help='Path to video file.')
25 | args = parser.parse_args()
26 |
27 |
28 | def choose_run_mode():
29 | global outFilePath
30 | if args.image:
31 | # Open the image file
32 | if not os.path.isfile(args.image):
33 | print("Input image file ", args.image, " doesn't exist")
34 | sys.exit(1)
35 | cap = cv.VideoCapture(args.image)
36 | outFilePath = str(outFilePath / (args.image[:-4] + '_out.jpg'))
37 | elif args.video:
38 | # Open the video file
39 | if not os.path.isfile(args.video):
40 | print("Input video file ", args.video, " doesn't exist")
41 | sys.exit(1)
42 | cap = cv.VideoCapture(args.video)
43 | outFilePath = str(outFilePath/(args.video[:-4]+'_out.mp4'))
44 | else:
45 | # Webcam input
46 | cap = cv.VideoCapture(0)
47 | outFilePath = str(outFilePath / 'webcam_out.mp4')
48 | return cap
49 |
50 |
51 | def load_pretrain_model():
52 | # 加载预训练后的POSE的caffe模型
53 | net = cv.dnn.readNetFromCaffe(proto_file, caffeModel=weights_file)
54 | print('POSE caffe model loaded successfully')
55 | # 调用GPU模块，但目前opencv仅支持intel GPU
56 | # tested with Intel GPUs only, or it will automatically switch to CPU
57 | net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
58 | net.setPreferableTarget(cv.dnn.DNN_TARGET_OPENCL)
59 | return net
60 |
61 |
62 | def show_fps():
63 | if not args.image:
64 | fps.update()
65 | fps.stop()
66 | fps_label = "FPS: {:.2f}".format(fps.fps())
67 | cv.putText(frame, fps_label, (0, origin_h - 25), cv.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 2)
68 |
69 |
70 | if __name__ == "__main__":
71 | cap = choose_run_mode()
72 | net = load_pretrain_model()
73 | fps = FPS().start()
74 | vid_writer = cv.VideoWriter(outFilePath, cv.VideoWriter_fourcc(*'mp4v'), 30,
75 | (round(cap.get(cv.CAP_PROP_FRAME_WIDTH)),
76 | round(cap.get(cv.CAP_PROP_FRAME_HEIGHT))))
77 | while cv.waitKey(1) < 0:
78 | hasFrame, frame = cap.read()
79 | if not hasFrame:
80 | print("Output file is stored as ", outFilePath)
81 | cv.waitKey(3000)
82 | break
83 |
84 | origin_h, origin_w = frame.shape[:2]
85 | blob = cv.dnn.blobFromImage(frame, 1.0 / 255, (input_width, input_height), 0, swapRB=False, crop=False)
86 | net.setInput(blob)
87 | detections = net.forward()
88 | H = detections.shape[2]
89 | W = detections.shape[3]
90 |
91 | # 存储关节点
92 | points = []
93 |
94 | for i in range(nPoints):
95 | probility_map = detections[0, i, :, :]
96 | #
97 | min_value, confidence, min_loc, point = cv.minMaxLoc(probility_map)
98 | #
99 | x = int(origin_w * (point[0] / W))
100 | y = int(origin_h * (point[1] / H))
101 | if confidence > threshold:
102 | cv.circle(frame, (x, y), 6, (255, 255, 0), -1, cv.FILLED)
103 | cv.putText(frame, "{}".format(i), (x, y-15), cv.FONT_HERSHEY_SIMPLEX, 0.4, (0, 0, 255), 1, cv.LINE_AA)
104 | points.append((x, y))
105 | else:
106 | points.append(None)
107 |
108 | # 画骨架
109 | for pair in POSE_PAIRS:
110 | A, B = pair[0], pair[1]
111 | if points[A] and points[B]:
112 | cv.line(frame, points[A], points[B], (0, 255, 255), 3, cv.LINE_AA)
113 | # cv.circle(frame, points[A], 8, (0, 0, 255), thickness=-1, lineType=cv.FILLED)
114 | # cv.circle(frame, points[B], 8, (0, 0, 255), thickness=-1, lineType=cv.FILLED)
115 |
116 | # Write the frame with the detection boxes
117 | if args.image:
118 | cv.imwrite(outFilePath, frame)
119 | else:
120 | vid_writer.write(frame)
121 |
122 | winName = 'Hand Skeleton from OpenPose'
123 | # cv.namedWindow(winName, cv.WINDOW_NORMAL)
124 | cv.imshow(winName, frame)
125 |
126 | if not args.image:
127 | vid_writer.release()
128 | cap.release()
129 | cv.destroyAllWindows()
--------------------------------------------------------------------------------
/hand-estimator-using-caffemodel/model/pose_deploy.prototxt:
--------------------------------------------------------------------------------
1 | input: "image"
2 | input_dim: 1 # Original: 2
3 | input_dim: 3 # It crashes if not left to 3
4 | input_dim: 1 # Original: 368
5 | input_dim: 1 # Original: 368
6 | layer {
7 | name: "conv1_1"
8 | type: "Convolution"
9 | bottom: "image"
10 | top: "conv1_1"
11 | param {
12 | lr_mult: 1.0
13 | decay_mult: 1
14 | }
15 | param {
16 | lr_mult: 2.0
17 | decay_mult: 0
18 | }
19 | convolution_param {
20 | num_output: 64
21 | pad: 1
22 | kernel_size: 3
23 | weight_filler {
24 | type: "xavier"
25 | }
26 | bias_filler {
27 | type: "constant"
28 | }
29 | dilation: 1
30 | }
31 | }
32 | layer {
33 | name: "relu1_1"
34 | type: "ReLU"
35 | bottom: "conv1_1"
36 | top: "conv1_1"
37 | }
38 | layer {
39 | name: "conv1_2"
40 | type: "Convolution"
41 | bottom: "conv1_1"
42 | top: "conv1_2"
43 | param {
44 | lr_mult: 1.0
45 | decay_mult: 1
46 | }
47 | param {
48 | lr_mult: 2.0
49 | decay_mult: 0
50 | }
51 | convolution_param {
52 | num_output: 64
53 | pad: 1
54 | kernel_size: 3
55 | weight_filler {
56 | type: "xavier"
57 | }
58 | bias_filler {
59 | type: "constant"
60 | }
61 | dilation: 1
62 | }
63 | }
64 | layer {
65 | name: "relu1_2"
66 | type: "ReLU"
67 | bottom: "conv1_2"
68 | top: "conv1_2"
69 | }
70 | layer {
71 | name: "pool1_stage1"
72 | type: "Pooling"
73 | bottom: "conv1_2"
74 | top: "pool1_stage1"
75 | pooling_param {
76 | pool: MAX
77 | kernel_size: 2
78 | stride: 2
79 | }
80 | }
81 | layer {
82 | name: "conv2_1"
83 | type: "Convolution"
84 | bottom: "pool1_stage1"
85 | top: "conv2_1"
86 | param {
87 | lr_mult: 1.0
88 | decay_mult: 1
89 | }
90 | param {
91 | lr_mult: 2.0
92 | decay_mult: 0
93 | }
94 | convolution_param {
95 | num_output: 128
96 | pad: 1
97 | kernel_size: 3
98 | weight_filler {
99 | type: "xavier"
100 | }
101 | bias_filler {
102 | type: "constant"
103 | }
104 | dilation: 1
105 | }
106 | }
107 | layer {
108 | name: "relu2_1"
109 | type: "ReLU"
110 | bottom: "conv2_1"
111 | top: "conv2_1"
112 | }
113 | layer {
114 | name: "conv2_2"
115 | type: "Convolution"
116 | bottom: "conv2_1"
117 | top: "conv2_2"
118 | param {
119 | lr_mult: 1.0
120 | decay_mult: 1
121 | }
122 | param {
123 | lr_mult: 2.0
124 | decay_mult: 0
125 | }
126 | convolution_param {
127 | num_output: 128
128 | pad: 1
129 | kernel_size: 3
130 | weight_filler {
131 | type: "xavier"
132 | }
133 | bias_filler {
134 | type: "constant"
135 | }
136 | dilation: 1
137 | }
138 | }
139 | layer {
140 | name: "relu2_2"
141 | type: "ReLU"
142 | bottom: "conv2_2"
143 | top: "conv2_2"
144 | }
145 | layer {
146 | name: "pool2_stage1"
147 | type: "Pooling"
148 | bottom: "conv2_2"
149 | top: "pool2_stage1"
150 | pooling_param {
151 | pool: MAX
152 | kernel_size: 2
153 | stride: 2
154 | }
155 | }
156 | layer {
157 | name: "conv3_1"
158 | type: "Convolution"
159 | bottom: "pool2_stage1"
160 | top: "conv3_1"
161 | param {
162 | lr_mult: 1.0
163 | decay_mult: 1
164 | }
165 | param {
166 | lr_mult: 2.0
167 | decay_mult: 0
168 | }
169 | convolution_param {
170 | num_output: 256
171 | pad: 1
172 | kernel_size: 3
173 | weight_filler {
174 | type: "xavier"
175 | }
176 | bias_filler {
177 | type: "constant"
178 | }
179 | dilation: 1
180 | }
181 | }
182 | layer {
183 | name: "relu3_1"
184 | type: "ReLU"
185 | bottom: "conv3_1"
186 | top: "conv3_1"
187 | }
188 | layer {
189 | name: "conv3_2"
190 | type: "Convolution"
191 | bottom: "conv3_1"
192 | top: "conv3_2"
193 | param {
194 | lr_mult: 1.0
195 | decay_mult: 1
196 | }
197 | param {
198 | lr_mult: 2.0
199 | decay_mult: 0
200 | }
201 | convolution_param {
202 | num_output: 256
203 | pad: 1
204 | kernel_size: 3
205 | weight_filler {
206 | type: "xavier"
207 | }
208 | bias_filler {
209 | type: "constant"
210 | }
211 | dilation: 1
212 | }
213 | }
214 | layer {
215 | name: "relu3_2"
216 | type: "ReLU"
217 | bottom: "conv3_2"
218 | top: "conv3_2"
219 | }
220 | layer {
221 | name: "conv3_3"
222 | type: "Convolution"
223 | bottom: "conv3_2"
224 | top: "conv3_3"
225 | param {
226 | lr_mult: 1.0
227 | decay_mult: 1
228 | }
229 | param {
230 | lr_mult: 2.0
231 | decay_mult: 0
232 | }
233 | convolution_param {
234 | num_output: 256
235 | pad: 1
236 | kernel_size: 3
237 | weight_filler {
238 | type: "xavier"
239 | }
240 | bias_filler {
241 | type: "constant"
242 | }
243 | dilation: 1
244 | }
245 | }
246 | layer {
247 | name: "relu3_3"
248 | type: "ReLU"
249 | bottom: "conv3_3"
250 | top: "conv3_3"
251 | }
252 | layer {
253 | name: "conv3_4"
254 | type: "Convolution"
255 | bottom: "conv3_3"
256 | top: "conv3_4"
257 | param {
258 | lr_mult: 1.0
259 | decay_mult: 1
260 | }
261 | param {
262 | lr_mult: 2.0
263 | decay_mult: 0
264 | }
265 | convolution_param {
266 | num_output: 256
267 | pad: 1
268 | kernel_size: 3
269 | weight_filler {
270 | type: "xavier"
271 | }
272 | bias_filler {
273 | type: "constant"
274 | }
275 | dilation: 1
276 | }
277 | }
278 | layer {
279 | name: "relu3_4"
280 | type: "ReLU"
281 | bottom: "conv3_4"
282 | top: "conv3_4"
283 | }
284 | layer {
285 | name: "pool3_stage1"
286 | type: "Pooling"
287 | bottom: "conv3_4"
288 | top: "pool3_stage1"
289 | pooling_param {
290 | pool: MAX
291 | kernel_size: 2
292 | stride: 2
293 | }
294 | }
295 | layer {
296 | name: "conv4_1"
297 | type: "Convolution"
298 | bottom: "pool3_stage1"
299 | top: "conv4_1"
300 | param {
301 | lr_mult: 1.0
302 | decay_mult: 1
303 | }
304 | param {
305 | lr_mult: 2.0
306 | decay_mult: 0
307 | }
308 | convolution_param {
309 | num_output: 512
310 | pad: 1
311 | kernel_size: 3
312 | weight_filler {
313 | type: "xavier"
314 | }
315 | bias_filler {
316 | type: "constant"
317 | }
318 | dilation: 1
319 | }
320 | }
321 | layer {
322 | name: "relu4_1"
323 | type: "ReLU"
324 | bottom: "conv4_1"
325 | top: "conv4_1"
326 | }
327 | layer {
328 | name: "conv4_2"
329 | type: "Convolution"
330 | bottom: "conv4_1"
331 | top: "conv4_2"
332 | param {
333 | lr_mult: 1.0
334 | decay_mult: 1
335 | }
336 | param {
337 | lr_mult: 2.0
338 | decay_mult: 0
339 | }
340 | convolution_param {
341 | num_output: 512
342 | pad: 1
343 | kernel_size: 3
344 | weight_filler {
345 | type: "xavier"
346 | }
347 | bias_filler {
348 | type: "constant"
349 | }
350 | dilation: 1
351 | }
352 | }
353 | layer {
354 | name: "relu4_2"
355 | type: "ReLU"
356 | bottom: "conv4_2"
357 | top: "conv4_2"
358 | }
359 | layer {
360 | name: "conv4_3"
361 | type: "Convolution"
362 | bottom: "conv4_2"
363 | top: "conv4_3"
364 | param {
365 | lr_mult: 1.0
366 | decay_mult: 1
367 | }
368 | param {
369 | lr_mult: 2.0
370 | decay_mult: 0
371 | }
372 | convolution_param {
373 | num_output: 512
374 | pad: 1
375 | kernel_size: 3
376 | weight_filler {
377 | type: "xavier"
378 | }
379 | bias_filler {
380 | type: "constant"
381 | }
382 | dilation: 1
383 | }
384 | }
385 | layer {
386 | name: "relu4_3"
387 | type: "ReLU"
388 | bottom: "conv4_3"
389 | top: "conv4_3"
390 | }
391 | layer {
392 | name: "conv4_4"
393 | type: "Convolution"
394 | bottom: "conv4_3"
395 | top: "conv4_4"
396 | param {
397 | lr_mult: 1.0
398 | decay_mult: 1
399 | }
400 | param {
401 | lr_mult: 2.0
402 | decay_mult: 0
403 | }
404 | convolution_param {
405 | num_output: 512
406 | pad: 1
407 | kernel_size: 3
408 | weight_filler {
409 | type: "xavier"
410 | }
411 | bias_filler {
412 | type: "constant"
413 | }
414 | dilation: 1
415 | }
416 | }
417 | layer {
418 | name: "relu4_4"
419 | type: "ReLU"
420 | bottom: "conv4_4"
421 | top: "conv4_4"
422 | }
423 | layer {
424 | name: "conv5_1"
425 | type: "Convolution"
426 | bottom: "conv4_4"
427 | top: "conv5_1"
428 | param {
429 | lr_mult: 1.0
430 | decay_mult: 1
431 | }
432 | param {
433 | lr_mult: 2.0
434 | decay_mult: 0
435 | }
436 | convolution_param {
437 | num_output: 512
438 | pad: 1
439 | kernel_size: 3
440 | weight_filler {
441 | type: "xavier"
442 | }
443 | bias_filler {
444 | type: "constant"
445 | }
446 | dilation: 1
447 | }
448 | }
449 | layer {
450 | name: "relu5_1"
451 | type: "ReLU"
452 | bottom: "conv5_1"
453 | top: "conv5_1"
454 | }
455 | layer {
456 | name: "conv5_2"
457 | type: "Convolution"
458 | bottom: "conv5_1"
459 | top: "conv5_2"
460 | param {
461 | lr_mult: 1.0
462 | decay_mult: 1
463 | }
464 | param {
465 | lr_mult: 2.0
466 | decay_mult: 0
467 | }
468 | convolution_param {
469 | num_output: 512
470 | pad: 1
471 | kernel_size: 3
472 | weight_filler {
473 | type: "xavier"
474 | }
475 | bias_filler {
476 | type: "constant"
477 | }
478 | dilation: 1
479 | }
480 | }
481 | layer {
482 | name: "relu5_2"
483 | type: "ReLU"
484 | bottom: "conv5_2"
485 | top: "conv5_2"
486 | }
487 | layer {
488 | name: "conv5_3_CPM"
489 | type: "Convolution"
490 | bottom: "conv5_2"
491 | top: "conv5_3_CPM"
492 | param {
493 | lr_mult: 1.0
494 | decay_mult: 1
495 | }
496 | param {
497 | lr_mult: 2.0
498 | decay_mult: 0
499 | }
500 | convolution_param {
501 | num_output: 128
502 | pad: 1
503 | kernel_size: 3
504 | weight_filler {
505 | type: "gaussian"
506 | std: 0.01
507 | }
508 | bias_filler {
509 | type: "constant"
510 | }
511 | dilation: 1
512 | }
513 | }
514 | layer {
515 | name: "relu5_4_stage1_3"
516 | type: "ReLU"
517 | bottom: "conv5_3_CPM"
518 | top: "conv5_3_CPM"
519 | }
520 | layer {
521 | name: "conv6_1_CPM"
522 | type: "Convolution"
523 | bottom: "conv5_3_CPM"
524 | top: "conv6_1_CPM"
525 | param {
526 | lr_mult: 1.0
527 | decay_mult: 1
528 | }
529 | param {
530 | lr_mult: 2.0
531 | decay_mult: 0
532 | }
533 | convolution_param {
534 | num_output: 512
535 | pad: 0
536 | kernel_size: 1
537 | weight_filler {
538 | type: "gaussian"
539 | std: 0.01
540 | }
541 | bias_filler {
542 | type: "constant"
543 | }
544 | dilation: 1
545 | }
546 | }
547 | layer {
548 | name: "relu6_4_stage1_1"
549 | type: "ReLU"
550 | bottom: "conv6_1_CPM"
551 | top: "conv6_1_CPM"
552 | }
553 | layer {
554 | name: "conv6_2_CPM"
555 | type: "Convolution"
556 | bottom: "conv6_1_CPM"
557 | top: "conv6_2_CPM"
558 | param {
559 | lr_mult: 1.0
560 | decay_mult: 1
561 | }
562 | param {
563 | lr_mult: 2.0
564 | decay_mult: 0
565 | }
566 | convolution_param {
567 | num_output: 22
568 | pad: 0
569 | kernel_size: 1
570 | weight_filler {
571 | type: "gaussian"
572 | std: 0.01
573 | }
574 | bias_filler {
575 | type: "constant"
576 | }
577 | dilation: 1
578 | }
579 | }
580 | layer {
581 | name: "concat_stage2"
582 | type: "Concat"
583 | bottom: "conv6_2_CPM"
584 | bottom: "conv5_3_CPM"
585 | top: "concat_stage2"
586 | concat_param {
587 | axis: 1
588 | }
589 | }
590 | layer {
591 | name: "Mconv1_stage2"
592 | type: "Convolution"
593 | bottom: "concat_stage2"
594 | top: "Mconv1_stage2"
595 | param {
596 | lr_mult: 4.0
597 | decay_mult: 1
598 | }
599 | param {
600 | lr_mult: 8.0
601 | decay_mult: 0
602 | }
603 | convolution_param {
604 | num_output: 128
605 | pad: 3
606 | kernel_size: 7
607 | weight_filler {
608 | type: "gaussian"
609 | std: 0.01
610 | }
611 | bias_filler {
612 | type: "constant"
613 | }
614 | dilation: 1
615 | }
616 | }
617 | layer {
618 | name: "Mrelu1_2_stage2_1"
619 | type: "ReLU"
620 | bottom: "Mconv1_stage2"
621 | top: "Mconv1_stage2"
622 | }
623 | layer {
624 | name: "Mconv2_stage2"
625 | type: "Convolution"
626 | bottom: "Mconv1_stage2"
627 | top: "Mconv2_stage2"
628 | param {
629 | lr_mult: 4.0
630 | decay_mult: 1
631 | }
632 | param {
633 | lr_mult: 8.0
634 | decay_mult: 0
635 | }
636 | convolution_param {
637 | num_output: 128
638 | pad: 3
639 | kernel_size: 7
640 | weight_filler {
641 | type: "gaussian"
642 | std: 0.01
643 | }
644 | bias_filler {
645 | type: "constant"
646 | }
647 | dilation: 1
648 | }
649 | }
650 | layer {
651 | name: "Mrelu1_3_stage2_2"
652 | type: "ReLU"
653 | bottom: "Mconv2_stage2"
654 | top: "Mconv2_stage2"
655 | }
656 | layer {
657 | name: "Mconv3_stage2"
658 | type: "Convolution"
659 | bottom: "Mconv2_stage2"
660 | top: "Mconv3_stage2"
661 | param {
662 | lr_mult: 4.0
663 | decay_mult: 1
664 | }
665 | param {
666 | lr_mult: 8.0
667 | decay_mult: 0
668 | }
669 | convolution_param {
670 | num_output: 128
671 | pad: 3
672 | kernel_size: 7
673 | weight_filler {
674 | type: "gaussian"
675 | std: 0.01
676 | }
677 | bias_filler {
678 | type: "constant"
679 | }
680 | dilation: 1
681 | }
682 | }
683 | layer {
684 | name: "Mrelu1_4_stage2_3"
685 | type: "ReLU"
686 | bottom: "Mconv3_stage2"
687 | top: "Mconv3_stage2"
688 | }
689 | layer {
690 | name: "Mconv4_stage2"
691 | type: "Convolution"
692 | bottom: "Mconv3_stage2"
693 | top: "Mconv4_stage2"
694 | param {
695 | lr_mult: 4.0
696 | decay_mult: 1
697 | }
698 | param {
699 | lr_mult: 8.0
700 | decay_mult: 0
701 | }
702 | convolution_param {
703 | num_output: 128
704 | pad: 3
705 | kernel_size: 7
706 | weight_filler {
707 | type: "gaussian"
708 | std: 0.01
709 | }
710 | bias_filler {
711 | type: "constant"
712 | }
713 | dilation: 1
714 | }
715 | }
716 | layer {
717 | name: "Mrelu1_5_stage2_4"
718 | type: "ReLU"
719 | bottom: "Mconv4_stage2"
720 | top: "Mconv4_stage2"
721 | }
722 | layer {
723 | name: "Mconv5_stage2"
724 | type: "Convolution"
725 | bottom: "Mconv4_stage2"
726 | top: "Mconv5_stage2"
727 | param {
728 | lr_mult: 4.0
729 | decay_mult: 1
730 | }
731 | param {
732 | lr_mult: 8.0
733 | decay_mult: 0
734 | }
735 | convolution_param {
736 | num_output: 128
737 | pad: 3
738 | kernel_size: 7
739 | weight_filler {
740 | type: "gaussian"
741 | std: 0.01
742 | }
743 | bias_filler {
744 | type: "constant"
745 | }
746 | dilation: 1
747 | }
748 | }
749 | layer {
750 | name: "Mrelu1_6_stage2_5"
751 | type: "ReLU"
752 | bottom: "Mconv5_stage2"
753 | top: "Mconv5_stage2"
754 | }
755 | layer {
756 | name: "Mconv6_stage2"
757 | type: "Convolution"
758 | bottom: "Mconv5_stage2"
759 | top: "Mconv6_stage2"
760 | param {
761 | lr_mult: 4.0
762 | decay_mult: 1
763 | }
764 | param {
765 | lr_mult: 8.0
766 | decay_mult: 0
767 | }
768 | convolution_param {
769 | num_output: 128
770 | pad: 0
771 | kernel_size: 1
772 | weight_filler {
773 | type: "gaussian"
774 | std: 0.01
775 | }
776 | bias_filler {
777 | type: "constant"
778 | }
779 | dilation: 1
780 | }
781 | }
782 | layer {
783 | name: "Mrelu1_7_stage2_6"
784 | type: "ReLU"
785 | bottom: "Mconv6_stage2"
786 | top: "Mconv6_stage2"
787 | }
788 | layer {
789 | name: "Mconv7_stage2"
790 | type: "Convolution"
791 | bottom: "Mconv6_stage2"
792 | top: "Mconv7_stage2"
793 | param {
794 | lr_mult: 4.0
795 | decay_mult: 1
796 | }
797 | param {
798 | lr_mult: 8.0
799 | decay_mult: 0
800 | }
801 | convolution_param {
802 | num_output: 22
803 | pad: 0
804 | kernel_size: 1
805 | weight_filler {
806 | type: "gaussian"
807 | std: 0.01
808 | }
809 | bias_filler {
810 | type: "constant"
811 | }
812 | dilation: 1
813 | }
814 | }
815 | layer {
816 | name: "concat_stage3"
817 | type: "Concat"
818 | bottom: "Mconv7_stage2"
819 | bottom: "conv5_3_CPM"
820 | top: "concat_stage3"
821 | concat_param {
822 | axis: 1
823 | }
824 | }
825 | layer {
826 | name: "Mconv1_stage3"
827 | type: "Convolution"
828 | bottom: "concat_stage3"
829 | top: "Mconv1_stage3"
830 | param {
831 | lr_mult: 4.0
832 | decay_mult: 1
833 | }
834 | param {
835 | lr_mult: 8.0
836 | decay_mult: 0
837 | }
838 | convolution_param {
839 | num_output: 128
840 | pad: 3
841 | kernel_size: 7
842 | weight_filler {
843 | type: "gaussian"
844 | std: 0.01
845 | }
846 | bias_filler {
847 | type: "constant"
848 | }
849 | dilation: 1
850 | }
851 | }
852 | layer {
853 | name: "Mrelu1_2_stage3_1"
854 | type: "ReLU"
855 | bottom: "Mconv1_stage3"
856 | top: "Mconv1_stage3"
857 | }
858 | layer {
859 | name: "Mconv2_stage3"
860 | type: "Convolution"
861 | bottom: "Mconv1_stage3"
862 | top: "Mconv2_stage3"
863 | param {
864 | lr_mult: 4.0
865 | decay_mult: 1
866 | }
867 | param {
868 | lr_mult: 8.0
869 | decay_mult: 0
870 | }
871 | convolution_param {
872 | num_output: 128
873 | pad: 3
874 | kernel_size: 7
875 | weight_filler {
876 | type: "gaussian"
877 | std: 0.01
878 | }
879 | bias_filler {
880 | type: "constant"
881 | }
882 | dilation: 1
883 | }
884 | }
885 | layer {
886 | name: "Mrelu1_3_stage3_2"
887 | type: "ReLU"
888 | bottom: "Mconv2_stage3"
889 | top: "Mconv2_stage3"
890 | }
891 | layer {
892 | name: "Mconv3_stage3"
893 | type: "Convolution"
894 | bottom: "Mconv2_stage3"
895 | top: "Mconv3_stage3"
896 | param {
897 | lr_mult: 4.0
898 | decay_mult: 1
899 | }
900 | param {
901 | lr_mult: 8.0
902 | decay_mult: 0
903 | }
904 | convolution_param {
905 | num_output: 128
906 | pad: 3
907 | kernel_size: 7
908 | weight_filler {
909 | type: "gaussian"
910 | std: 0.01
911 | }
912 | bias_filler {
913 | type: "constant"
914 | }
915 | dilation: 1
916 | }
917 | }
918 | layer {
919 | name: "Mrelu1_4_stage3_3"
920 | type: "ReLU"
921 | bottom: "Mconv3_stage3"
922 | top: "Mconv3_stage3"
923 | }
924 | layer {
925 | name: "Mconv4_stage3"
926 | type: "Convolution"
927 | bottom: "Mconv3_stage3"
928 | top: "Mconv4_stage3"
929 | param {
930 | lr_mult: 4.0
931 | decay_mult: 1
932 | }
933 | param {
934 | lr_mult: 8.0
935 | decay_mult: 0
936 | }
937 | convolution_param {
938 | num_output: 128
939 | pad: 3
940 | kernel_size: 7
941 | weight_filler {
942 | type: "gaussian"
943 | std: 0.01
944 | }
945 | bias_filler {
946 | type: "constant"
947 | }
948 | dilation: 1
949 | }
950 | }
951 | layer {
952 | name: "Mrelu1_5_stage3_4"
953 | type: "ReLU"
954 | bottom: "Mconv4_stage3"
955 | top: "Mconv4_stage3"
956 | }
957 | layer {
958 | name: "Mconv5_stage3"
959 | type: "Convolution"
960 | bottom: "Mconv4_stage3"
961 | top: "Mconv5_stage3"
962 | param {
963 | lr_mult: 4.0
964 | decay_mult: 1
965 | }
966 | param {
967 | lr_mult: 8.0
968 | decay_mult: 0
969 | }
970 | convolution_param {
971 | num_output: 128
972 | pad: 3
973 | kernel_size: 7
974 | weight_filler {
975 | type: "gaussian"
976 | std: 0.01
977 | }
978 | bias_filler {
979 | type: "constant"
980 | }
981 | dilation: 1
982 | }
983 | }
984 | layer {
985 | name: "Mrelu1_6_stage3_5"
986 | type: "ReLU"
987 | bottom: "Mconv5_stage3"
988 | top: "Mconv5_stage3"
989 | }
990 | layer {
991 | name: "Mconv6_stage3"
992 | type: "Convolution"
993 | bottom: "Mconv5_stage3"
994 | top: "Mconv6_stage3"
995 | param {
996 | lr_mult: 4.0
997 | decay_mult: 1
998 | }
999 | param {
1000 | lr_mult: 8.0
1001 | decay_mult: 0
1002 | }
1003 | convolution_param {
1004 | num_output: 128
1005 | pad: 0
1006 | kernel_size: 1
1007 | weight_filler {
1008 | type: "gaussian"
1009 | std: 0.01
1010 | }
1011 | bias_filler {
1012 | type: "constant"
1013 | }
1014 | dilation: 1
1015 | }
1016 | }
1017 | layer {
1018 | name: "Mrelu1_7_stage3_6"
1019 | type: "ReLU"
1020 | bottom: "Mconv6_stage3"
1021 | top: "Mconv6_stage3"
1022 | }
1023 | layer {
1024 | name: "Mconv7_stage3"
1025 | type: "Convolution"
1026 | bottom: "Mconv6_stage3"
1027 | top: "Mconv7_stage3"
1028 | param {
1029 | lr_mult: 4.0
1030 | decay_mult: 1
1031 | }
1032 | param {
1033 | lr_mult: 8.0
1034 | decay_mult: 0
1035 | }
1036 | convolution_param {
1037 | num_output: 22
1038 | pad: 0
1039 | kernel_size: 1
1040 | weight_filler {
1041 | type: "gaussian"
1042 | std: 0.01
1043 | }
1044 | bias_filler {
1045 | type: "constant"
1046 | }
1047 | dilation: 1
1048 | }
1049 | }
1050 | layer {
1051 | name: "concat_stage4"
1052 | type: "Concat"
1053 | bottom: "Mconv7_stage3"
1054 | bottom: "conv5_3_CPM"
1055 | top: "concat_stage4"
1056 | concat_param {
1057 | axis: 1
1058 | }
1059 | }
1060 | layer {
1061 | name: "Mconv1_stage4"
1062 | type: "Convolution"
1063 | bottom: "concat_stage4"
1064 | top: "Mconv1_stage4"
1065 | param {
1066 | lr_mult: 4.0
1067 | decay_mult: 1
1068 | }
1069 | param {
1070 | lr_mult: 8.0
1071 | decay_mult: 0
1072 | }
1073 | convolution_param {
1074 | num_output: 128
1075 | pad: 3
1076 | kernel_size: 7
1077 | weight_filler {
1078 | type: "gaussian"
1079 | std: 0.01
1080 | }
1081 | bias_filler {
1082 | type: "constant"
1083 | }
1084 | dilation: 1
1085 | }
1086 | }
1087 | layer {
1088 | name: "Mrelu1_2_stage4_1"
1089 | type: "ReLU"
1090 | bottom: "Mconv1_stage4"
1091 | top: "Mconv1_stage4"
1092 | }
1093 | layer {
1094 | name: "Mconv2_stage4"
1095 | type: "Convolution"
1096 | bottom: "Mconv1_stage4"
1097 | top: "Mconv2_stage4"
1098 | param {
1099 | lr_mult: 4.0
1100 | decay_mult: 1
1101 | }
1102 | param {
1103 | lr_mult: 8.0
1104 | decay_mult: 0
1105 | }
1106 | convolution_param {
1107 | num_output: 128
1108 | pad: 3
1109 | kernel_size: 7
1110 | weight_filler {
1111 | type: "gaussian"
1112 | std: 0.01
1113 | }
1114 | bias_filler {
1115 | type: "constant"
1116 | }
1117 | dilation: 1
1118 | }
1119 | }
1120 | layer {
1121 | name: "Mrelu1_3_stage4_2"
1122 | type: "ReLU"
1123 | bottom: "Mconv2_stage4"
1124 | top: "Mconv2_stage4"
1125 | }
1126 | layer {
1127 | name: "Mconv3_stage4"
1128 | type: "Convolution"
1129 | bottom: "Mconv2_stage4"
1130 | top: "Mconv3_stage4"
1131 | param {
1132 | lr_mult: 4.0
1133 | decay_mult: 1
1134 | }
1135 | param {
1136 | lr_mult: 8.0
1137 | decay_mult: 0
1138 | }
1139 | convolution_param {
1140 | num_output: 128
1141 | pad: 3
1142 | kernel_size: 7
1143 | weight_filler {
1144 | type: "gaussian"
1145 | std: 0.01
1146 | }
1147 | bias_filler {
1148 | type: "constant"
1149 | }
1150 | dilation: 1
1151 | }
1152 | }
1153 | layer {
1154 | name: "Mrelu1_4_stage4_3"
1155 | type: "ReLU"
1156 | bottom: "Mconv3_stage4"
1157 | top: "Mconv3_stage4"
1158 | }
1159 | layer {
1160 | name: "Mconv4_stage4"
1161 | type: "Convolution"
1162 | bottom: "Mconv3_stage4"
1163 | top: "Mconv4_stage4"
1164 | param {
1165 | lr_mult: 4.0
1166 | decay_mult: 1
1167 | }
1168 | param {
1169 | lr_mult: 8.0
1170 | decay_mult: 0
1171 | }
1172 | convolution_param {
1173 | num_output: 128
1174 | pad: 3
1175 | kernel_size: 7
1176 | weight_filler {
1177 | type: "gaussian"
1178 | std: 0.01
1179 | }
1180 | bias_filler {
1181 | type: "constant"
1182 | }
1183 | dilation: 1
1184 | }
1185 | }
1186 | layer {
1187 | name: "Mrelu1_5_stage4_4"
1188 | type: "ReLU"
1189 | bottom: "Mconv4_stage4"
1190 | top: "Mconv4_stage4"
1191 | }
1192 | layer {
1193 | name: "Mconv5_stage4"
1194 | type: "Convolution"
1195 | bottom: "Mconv4_stage4"
1196 | top: "Mconv5_stage4"
1197 | param {
1198 | lr_mult: 4.0
1199 | decay_mult: 1
1200 | }
1201 | param {
1202 | lr_mult: 8.0
1203 | decay_mult: 0
1204 | }
1205 | convolution_param {
1206 | num_output: 128
1207 | pad: 3
1208 | kernel_size: 7
1209 | weight_filler {
1210 | type: "gaussian"
1211 | std: 0.01
1212 | }
1213 | bias_filler {
1214 | type: "constant"
1215 | }
1216 | dilation: 1
1217 | }
1218 | }
1219 | layer {
1220 | name: "Mrelu1_6_stage4_5"
1221 | type: "ReLU"
1222 | bottom: "Mconv5_stage4"
1223 | top: "Mconv5_stage4"
1224 | }
1225 | layer {
1226 | name: "Mconv6_stage4"
1227 | type: "Convolution"
1228 | bottom: "Mconv5_stage4"
1229 | top: "Mconv6_stage4"
1230 | param {
1231 | lr_mult: 4.0
1232 | decay_mult: 1
1233 | }
1234 | param {
1235 | lr_mult: 8.0
1236 | decay_mult: 0
1237 | }
1238 | convolution_param {
1239 | num_output: 128
1240 | pad: 0
1241 | kernel_size: 1
1242 | weight_filler {
1243 | type: "gaussian"
1244 | std: 0.01
1245 | }
1246 | bias_filler {
1247 | type: "constant"
1248 | }
1249 | dilation: 1
1250 | }
1251 | }
1252 | layer {
1253 | name: "Mrelu1_7_stage4_6"
1254 | type: "ReLU"
1255 | bottom: "Mconv6_stage4"
1256 | top: "Mconv6_stage4"
1257 | }
1258 | layer {
1259 | name: "Mconv7_stage4"
1260 | type: "Convolution"
1261 | bottom: "Mconv6_stage4"
1262 | top: "Mconv7_stage4"
1263 | param {
1264 | lr_mult: 4.0
1265 | decay_mult: 1
1266 | }
1267 | param {
1268 | lr_mult: 8.0
1269 | decay_mult: 0
1270 | }
1271 | convolution_param {
1272 | num_output: 22
1273 | pad: 0
1274 | kernel_size: 1
1275 | weight_filler {
1276 | type: "gaussian"
1277 | std: 0.01
1278 | }
1279 | bias_filler {
1280 | type: "constant"
1281 | }
1282 | dilation: 1
1283 | }
1284 | }
1285 | layer {
1286 | name: "concat_stage5"
1287 | type: "Concat"
1288 | bottom: "Mconv7_stage4"
1289 | bottom: "conv5_3_CPM"
1290 | top: "concat_stage5"
1291 | concat_param {
1292 | axis: 1
1293 | }
1294 | }
1295 | layer {
1296 | name: "Mconv1_stage5"
1297 | type: "Convolution"
1298 | bottom: "concat_stage5"
1299 | top: "Mconv1_stage5"
1300 | param {
1301 | lr_mult: 4.0
1302 | decay_mult: 1
1303 | }
1304 | param {
1305 | lr_mult: 8.0
1306 | decay_mult: 0
1307 | }
1308 | convolution_param {
1309 | num_output: 128
1310 | pad: 3
1311 | kernel_size: 7
1312 | weight_filler {
1313 | type: "gaussian"
1314 | std: 0.01
1315 | }
1316 | bias_filler {
1317 | type: "constant"
1318 | }
1319 | dilation: 1
1320 | }
1321 | }
1322 | layer {
1323 | name: "Mrelu1_2_stage5_1"
1324 | type: "ReLU"
1325 | bottom: "Mconv1_stage5"
1326 | top: "Mconv1_stage5"
1327 | }
1328 | layer {
1329 | name: "Mconv2_stage5"
1330 | type: "Convolution"
1331 | bottom: "Mconv1_stage5"
1332 | top: "Mconv2_stage5"
1333 | param {
1334 | lr_mult: 4.0
1335 | decay_mult: 1
1336 | }
1337 | param {
1338 | lr_mult: 8.0
1339 | decay_mult: 0
1340 | }
1341 | convolution_param {
1342 | num_output: 128
1343 | pad: 3
1344 | kernel_size: 7
1345 | weight_filler {
1346 | type: "gaussian"
1347 | std: 0.01
1348 | }
1349 | bias_filler {
1350 | type: "constant"
1351 | }
1352 | dilation: 1
1353 | }
1354 | }
1355 | layer {
1356 | name: "Mrelu1_3_stage5_2"
1357 | type: "ReLU"
1358 | bottom: "Mconv2_stage5"
1359 | top: "Mconv2_stage5"
1360 | }
1361 | layer {
1362 | name: "Mconv3_stage5"
1363 | type: "Convolution"
1364 | bottom: "Mconv2_stage5"
1365 | top: "Mconv3_stage5"
1366 | param {
1367 | lr_mult: 4.0
1368 | decay_mult: 1
1369 | }
1370 | param {
1371 | lr_mult: 8.0
1372 | decay_mult: 0
1373 | }
1374 | convolution_param {
1375 | num_output: 128
1376 | pad: 3
1377 | kernel_size: 7
1378 | weight_filler {
1379 | type: "gaussian"
1380 | std: 0.01
1381 | }
1382 | bias_filler {
1383 | type: "constant"
1384 | }
1385 | dilation: 1
1386 | }
1387 | }
1388 | layer {
1389 | name: "Mrelu1_4_stage5_3"
1390 | type: "ReLU"
1391 | bottom: "Mconv3_stage5"
1392 | top: "Mconv3_stage5"
1393 | }
1394 | layer {
1395 | name: "Mconv4_stage5"
1396 | type: "Convolution"
1397 | bottom: "Mconv3_stage5"
1398 | top: "Mconv4_stage5"
1399 | param {
1400 | lr_mult: 4.0
1401 | decay_mult: 1
1402 | }
1403 | param {
1404 | lr_mult: 8.0
1405 | decay_mult: 0
1406 | }
1407 | convolution_param {
1408 | num_output: 128
1409 | pad: 3
1410 | kernel_size: 7
1411 | weight_filler {
1412 | type: "gaussian"
1413 | std: 0.01
1414 | }
1415 | bias_filler {
1416 | type: "constant"
1417 | }
1418 | dilation: 1
1419 | }
1420 | }
1421 | layer {
1422 | name: "Mrelu1_5_stage5_4"
1423 | type: "ReLU"
1424 | bottom: "Mconv4_stage5"
1425 | top: "Mconv4_stage5"
1426 | }
1427 | layer {
1428 | name: "Mconv5_stage5"
1429 | type: "Convolution"
1430 | bottom: "Mconv4_stage5"
1431 | top: "Mconv5_stage5"
1432 | param {
1433 | lr_mult: 4.0
1434 | decay_mult: 1
1435 | }
1436 | param {
1437 | lr_mult: 8.0
1438 | decay_mult: 0
1439 | }
1440 | convolution_param {
1441 | num_output: 128
1442 | pad: 3
1443 | kernel_size: 7
1444 | weight_filler {
1445 | type: "gaussian"
1446 | std: 0.01
1447 | }
1448 | bias_filler {
1449 | type: "constant"
1450 | }
1451 | dilation: 1
1452 | }
1453 | }
1454 | layer {
1455 | name: "Mrelu1_6_stage5_5"
1456 | type: "ReLU"
1457 | bottom: "Mconv5_stage5"
1458 | top: "Mconv5_stage5"
1459 | }
1460 | layer {
1461 | name: "Mconv6_stage5"
1462 | type: "Convolution"
1463 | bottom: "Mconv5_stage5"
1464 | top: "Mconv6_stage5"
1465 | param {
1466 | lr_mult: 4.0
1467 | decay_mult: 1
1468 | }
1469 | param {
1470 | lr_mult: 8.0
1471 | decay_mult: 0
1472 | }
1473 | convolution_param {
1474 | num_output: 128
1475 | pad: 0
1476 | kernel_size: 1
1477 | weight_filler {
1478 | type: "gaussian"
1479 | std: 0.01
1480 | }
1481 | bias_filler {
1482 | type: "constant"
1483 | }
1484 | dilation: 1
1485 | }
1486 | }
1487 | layer {
1488 | name: "Mrelu1_7_stage5_6"
1489 | type: "ReLU"
1490 | bottom: "Mconv6_stage5"
1491 | top: "Mconv6_stage5"
1492 | }
1493 | layer {
1494 | name: "Mconv7_stage5"
1495 | type: "Convolution"
1496 | bottom: "Mconv6_stage5"
1497 | top: "Mconv7_stage5"
1498 | param {
1499 | lr_mult: 4.0
1500 | decay_mult: 1
1501 | }
1502 | param {
1503 | lr_mult: 8.0
1504 | decay_mult: 0
1505 | }
1506 | convolution_param {
1507 | num_output: 22
1508 | pad: 0
1509 | kernel_size: 1
1510 | weight_filler {
1511 | type: "gaussian"
1512 | std: 0.01
1513 | }
1514 | bias_filler {
1515 | type: "constant"
1516 | }
1517 | dilation: 1
1518 | }
1519 | }
1520 | layer {
1521 | name: "concat_stage6"
1522 | type: "Concat"
1523 | bottom: "Mconv7_stage5"
1524 | bottom: "conv5_3_CPM"
1525 | top: "concat_stage6"
1526 | concat_param {
1527 | axis: 1
1528 | }
1529 | }
1530 | layer {
1531 | name: "Mconv1_stage6"
1532 | type: "Convolution"
1533 | bottom: "concat_stage6"
1534 | top: "Mconv1_stage6"
1535 | param {
1536 | lr_mult: 4.0
1537 | decay_mult: 1
1538 | }
1539 | param {
1540 | lr_mult: 8.0
1541 | decay_mult: 0
1542 | }
1543 | convolution_param {
1544 | num_output: 128
1545 | pad: 3
1546 | kernel_size: 7
1547 | weight_filler {
1548 | type: "gaussian"
1549 | std: 0.01
1550 | }
1551 | bias_filler {
1552 | type: "constant"
1553 | }
1554 | dilation: 1
1555 | }
1556 | }
1557 | layer {
1558 | name: "Mrelu1_2_stage6_1"
1559 | type: "ReLU"
1560 | bottom: "Mconv1_stage6"
1561 | top: "Mconv1_stage6"
1562 | }
1563 | layer {
1564 | name: "Mconv2_stage6"
1565 | type: "Convolution"
1566 | bottom: "Mconv1_stage6"
1567 | top: "Mconv2_stage6"
1568 | param {
1569 | lr_mult: 4.0
1570 | decay_mult: 1
1571 | }
1572 | param {
1573 | lr_mult: 8.0
1574 | decay_mult: 0
1575 | }
1576 | convolution_param {
1577 | num_output: 128
1578 | pad: 3
1579 | kernel_size: 7
1580 | weight_filler {
1581 | type: "gaussian"
1582 | std: 0.01
1583 | }
1584 | bias_filler {
1585 | type: "constant"
1586 | }
1587 | dilation: 1
1588 | }
1589 | }
1590 | layer {
1591 | name: "Mrelu1_3_stage6_2"
1592 | type: "ReLU"
1593 | bottom: "Mconv2_stage6"
1594 | top: "Mconv2_stage6"
1595 | }
1596 | layer {
1597 | name: "Mconv3_stage6"
1598 | type: "Convolution"
1599 | bottom: "Mconv2_stage6"
1600 | top: "Mconv3_stage6"
1601 | param {
1602 | lr_mult: 4.0
1603 | decay_mult: 1
1604 | }
1605 | param {
1606 | lr_mult: 8.0
1607 | decay_mult: 0
1608 | }
1609 | convolution_param {
1610 | num_output: 128
1611 | pad: 3
1612 | kernel_size: 7
1613 | weight_filler {
1614 | type: "gaussian"
1615 | std: 0.01
1616 | }
1617 | bias_filler {
1618 | type: "constant"
1619 | }
1620 | dilation: 1
1621 | }
1622 | }
1623 | layer {
1624 | name: "Mrelu1_4_stage6_3"
1625 | type: "ReLU"
1626 | bottom: "Mconv3_stage6"
1627 | top: "Mconv3_stage6"
1628 | }
1629 | layer {
1630 | name: "Mconv4_stage6"
1631 | type: "Convolution"
1632 | bottom: "Mconv3_stage6"
1633 | top: "Mconv4_stage6"
1634 | param {
1635 | lr_mult: 4.0
1636 | decay_mult: 1
1637 | }
1638 | param {
1639 | lr_mult: 8.0
1640 | decay_mult: 0
1641 | }
1642 | convolution_param {
1643 | num_output: 128
1644 | pad: 3
1645 | kernel_size: 7
1646 | weight_filler {
1647 | type: "gaussian"
1648 | std: 0.01
1649 | }
1650 | bias_filler {
1651 | type: "constant"
1652 | }
1653 | dilation: 1
1654 | }
1655 | }
1656 | layer {
1657 | name: "Mrelu1_5_stage6_4"
1658 | type: "ReLU"
1659 | bottom: "Mconv4_stage6"
1660 | top: "Mconv4_stage6"
1661 | }
1662 | layer {
1663 | name: "Mconv5_stage6"
1664 | type: "Convolution"
1665 | bottom: "Mconv4_stage6"
1666 | top: "Mconv5_stage6"
1667 | param {
1668 | lr_mult: 4.0
1669 | decay_mult: 1
1670 | }
1671 | param {
1672 | lr_mult: 8.0
1673 | decay_mult: 0
1674 | }
1675 | convolution_param {
1676 | num_output: 128
1677 | pad: 3
1678 | kernel_size: 7
1679 | weight_filler {
1680 | type: "gaussian"
1681 | std: 0.01
1682 | }
1683 | bias_filler {
1684 | type: "constant"
1685 | }
1686 | dilation: 1
1687 | }
1688 | }
1689 | layer {
1690 | name: "Mrelu1_6_stage6_5"
1691 | type: "ReLU"
1692 | bottom: "Mconv5_stage6"
1693 | top: "Mconv5_stage6"
1694 | }
1695 | layer {
1696 | name: "Mconv6_stage6"
1697 | type: "Convolution"
1698 | bottom: "Mconv5_stage6"
1699 | top: "Mconv6_stage6"
1700 | param {
1701 | lr_mult: 4.0
1702 | decay_mult: 1
1703 | }
1704 | param {
1705 | lr_mult: 8.0
1706 | decay_mult: 0
1707 | }
1708 | convolution_param {
1709 | num_output: 128
1710 | pad: 0
1711 | kernel_size: 1
1712 | weight_filler {
1713 | type: "gaussian"
1714 | std: 0.01
1715 | }
1716 | bias_filler {
1717 | type: "constant"
1718 | }
1719 | dilation: 1
1720 | }
1721 | }
1722 | layer {
1723 | name: "Mrelu1_7_stage6_6"
1724 | type: "ReLU"
1725 | bottom: "Mconv6_stage6"
1726 | top: "Mconv6_stage6"
1727 | }
1728 | layer {
1729 | name: "Mconv7_stage6"
1730 | type: "Convolution"
1731 | bottom: "Mconv6_stage6"
1732 | # top: "Mconv7_stage6"
1733 | top: "net_output"
1734 | param {
1735 | lr_mult: 4.0
1736 | decay_mult: 1
1737 | }
1738 | param {
1739 | lr_mult: 8.0
1740 | decay_mult: 0
1741 | }
1742 | convolution_param {
1743 | num_output: 22
1744 | pad: 0
1745 | kernel_size: 1
1746 | weight_filler {
1747 | type: "gaussian"
1748 | std: 0.01
1749 | }
1750 | bias_filler {
1751 | type: "constant"
1752 | }
1753 | dilation: 1
1754 | }
1755 | }
1756 |
1757 |
--------------------------------------------------------------------------------
/hand-estimator-using-caffemodel/test_out/README.md:
--------------------------------------------------------------------------------
1 | This directory is for saving the after-processing outputs.
2 |
--------------------------------------------------------------------------------
/hand-estimator-using-caffemodel/test_out/hand_out.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/hand-estimator-using-caffemodel/test_out/hand_out.jpg
--------------------------------------------------------------------------------
/hand-estimator-using-caffemodel/test_out/shake_out.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/hand-estimator-using-caffemodel/test_out/shake_out.jpg
--------------------------------------------------------------------------------
/pose-estimator-tensorflow/README.md:
--------------------------------------------------------------------------------
1 |
2 | ## Usage Examples :
3 | ***for `openpose_tf.py`***
4 | 1. Creat a file names `graph_model_coco`, and palce the pretrained model here;
5 | 2. Put the test file (image or video) under the same directory of `openpose_tf.py`.
6 | 3. Commmand line input:
7 | - `python openpose_tf.py --image=test.jpg`
8 | - `python openpose_tf.py --video=test.mp4`
9 | - if no argument provided, it starts the webcam.
10 |
11 | ***for `openpose_skeleton_sequence_drawer.py`***
12 | 1. Same as the upper;
13 | 2. Commmand line input:
14 | - `python openpose_skeleton_sequence_drawer.py`
15 |
16 | ## Pretrained models intro
17 | - **graph_opt.pb**: training with the VGG net, as same as the CMU providing caffemodel, ***more accurate but slower***
18 | - **graph_opt_mobile.pb**: training with the Mobilenet, much smaller than the origin VGG, ***faster but less accurate***
19 |
20 | ## Acknowledgement
21 | Thanks [ildoonet](https://github.com/ildoonet) and his awesome work [tf-pose-estimation](https://github.com/ildoonet/tf-pose-estimation), the graph weight files are collected there.
22 |
--------------------------------------------------------------------------------
/pose-estimator-tensorflow/openpose_skeleton_sequence_drawer.py:
--------------------------------------------------------------------------------
1 | import cv2 as cv
2 | import numpy as np
3 | import os
4 | import math
5 | import argparse
6 | import sys
7 | import itertools
8 | import tensorflow as tf
9 | from enum import Enum
10 | from collections import namedtuple
11 | from scipy.ndimage import maximum_filter, gaussian_filter
12 | from imutils.video import FPS
13 | from pathlib import Path
14 | import time
15 |
16 |
17 | class Human:
18 | """
19 | body_parts: list of BodyPart
20 | """
21 | __slots__ = ('body_parts', 'pairs', 'uidx_list')
22 |
23 | def __init__(self, pairs):
24 | self.pairs = []
25 | self.uidx_list = set()
26 | self.body_parts = {}
27 | for pair in pairs:
28 | self.add_pair(pair)
29 |
30 | @staticmethod
31 | def _get_uidx(part_idx, idx):
32 | return '%d-%d' % (part_idx, idx)
33 |
34 | def add_pair(self, pair):
35 | self.pairs.append(pair)
36 | self.body_parts[pair.part_idx1] = BodyPart(Human._get_uidx(pair.part_idx1, pair.idx1),
37 | pair.part_idx1,
38 | pair.coord1[0], pair.coord1[1], pair.score)
39 | self.body_parts[pair.part_idx2] = BodyPart(Human._get_uidx(pair.part_idx2, pair.idx2),
40 | pair.part_idx2,
41 | pair.coord2[0], pair.coord2[1], pair.score)
42 | self.uidx_list.add(Human._get_uidx(pair.part_idx1, pair.idx1))
43 | self.uidx_list.add(Human._get_uidx(pair.part_idx2, pair.idx2))
44 |
45 | def is_connected(self, other):
46 | return len(self.uidx_list & other.uidx_list) > 0
47 |
48 | def merge(self, other):
49 | for pair in other.pairs:
50 | self.add_pair(pair)
51 |
52 | def part_count(self):
53 | return len(self.body_parts.keys())
54 |
55 | def get_max_score(self):
56 | return max([x.score for _, x in self.body_parts.items()])
57 |
58 | def __str__(self):
59 | return ' '.join([str(x) for x in self.body_parts.values()])
60 |
61 |
62 | class BodyPart:
63 | """
64 | part_idx : part index(eg. 0 for nose)
65 | x, y: coordinate of body part
66 | score : confidence score
67 | """
68 | __slots__ = ('uidx', 'part_idx', 'x', 'y', 'score')
69 |
70 | def __init__(self, uidx, part_idx, x, y, score):
71 | self.uidx = uidx
72 | self.part_idx = part_idx
73 | self.x, self.y = x, y
74 | self.score = score
75 |
76 |
77 | class TfPoseEstimator:
78 | ENSEMBLE = 'addup' # average, addup
79 |
80 | def __init__(self, graph_path, target_size=(368, 368)):
81 | self.target_size = target_size
82 |
83 | # load graph
84 | with tf.gfile.GFile(graph_path, 'rb') as f:
85 | graph_def = tf.GraphDef()
86 | graph_def.ParseFromString(f.read())
87 |
88 | self.graph = tf.get_default_graph()
89 | tf.import_graph_def(graph_def, name='TfPoseEstimator')
90 | self.persistent_sess = tf.Session(graph=self.graph)
91 |
92 | self.tensor_image = self.graph.get_tensor_by_name('TfPoseEstimator/image:0')
93 | self.tensor_output = self.graph.get_tensor_by_name('TfPoseEstimator/Openpose/concat_stage7:0')
94 | self.heatMat = self.pafMat = None
95 |
96 | @staticmethod
97 | def draw_pose_bbox(npimg, humans, imgcopy=False):
98 | if imgcopy:
99 | npimg = np.copy(npimg)
100 | image_h, image_w = npimg.shape[:2]
101 | joints, bboxes, xcenter = [], [], []
102 | for human in humans:
103 | xs, ys, centers = [], [], {}
104 | # 将所有关节点绘制到图像上
105 | for i in range(CocoPart.Background.value):
106 | if i not in human.body_parts.keys():
107 | continue
108 |
109 | body_part = human.body_parts[i]
110 | center = (int(body_part.x * image_w + 0.5), int(body_part.y * image_h + 0.5))
111 |
112 | centers[i] = center
113 | xs.append(center[0])
114 | ys.append(center[1])
115 | # 绘制关节点
116 | cv.circle(npimg, center, 3, CocoColors[i], thickness=3, lineType=8, shift=0)
117 | # 将属于同一人的关节点按照各个部位相连
118 | for pair_order, pair in enumerate(CocoPairsRender):
119 | if pair[0] not in human.body_parts.keys() or pair[1] not in human.body_parts.keys():
120 | continue
121 | cv.line(npimg, centers[pair[0]], centers[pair[1]], CocoColors[pair_order], 3, cv.LINE_AA)
122 | # 根据每个人的关节点信息生成ROI区域
123 | xmin = float(min(xs) / image_w)
124 | ymin = float(min(ys) / image_h)
125 | xmax = float(max(xs) / image_w)
126 | ymax = float(max(ys) / image_h)
127 | bboxes.append([xmin, ymin, xmax, ymax, 0.9999])
128 | joints.append(centers)
129 | if 1 in centers:
130 | xcenter.append(centers[1][0])
131 |
132 | # draw bounding_boxes
133 | # x_start, x_end= int(xmin*image_w) - 20, int(xmax*image_w) + 20
134 | # y_start, y_end =int(ymin*image_h) - 15, int(ymax*image_h) + 15
135 | # cv.rectangle(npimg, (x_start, y_start), (x_end, y_end), [0, 250, 0], 3)
136 | return npimg, joints, bboxes, xcenter
137 |
138 | @staticmethod
139 | def draw_pose_sequence(frame, humans, cnt):
140 | # global back_ground
141 | image_h, image_w = frame.shape[:2]
142 | for human in humans:
143 | xs, ys, centers = [], [], {}
144 | # 将所有关节点绘制到图像上
145 | joints_num = CocoPart.Background.value
146 | for i in range(joints_num):
147 | if i not in human.body_parts.keys():
148 | continue
149 |
150 | body_part = human.body_parts[i]
151 | center = (int(body_part.x * image_w + 0.5),
152 | int(body_part.y * image_h + 0.5))
153 |
154 | pos_move = cnt * back_ground_w / 500
155 | center = (int(center[0] + pos_move / 2),
156 | int(center[1] + back_ground_h / 1.5 - pos_move))
157 |
158 | centers[i] = center
159 | xs.append(center[0])
160 | ys.append(center[1])
161 | # 绘制关节点
162 | cv.circle(back_ground, centers[i], 3, CocoColors[i], thickness=3, lineType=8, shift=0)
163 |
164 | # 将属于同一人的关节点按照各个部位相连
165 | for pair_order, pair in enumerate(CocoPairsRender):
166 | if pair[0] not in human.body_parts.keys() or pair[1] not in human.body_parts.keys():
167 | continue
168 | cv.line(back_ground, centers[pair[0]], centers[pair[1]], CocoColors[pair_order], 3, cv.LINE_AA)
169 |
170 | @staticmethod
171 | def draw_pose_sequence_reverse(frame, humans, cnt):
172 | image_h, image_w = frame.shape[:2]
173 | for human in humans:
174 | xs, ys, centers = [], [], {}
175 | # 将所有关节点绘制到图像上
176 | joints_num = CocoPart.Background.value
177 | for i in range(joints_num):
178 | if i not in human.body_parts.keys():
179 | continue
180 |
181 | body_part = human.body_parts[i]
182 | center = (int(body_part.x * image_w + 0.5),
183 | int(body_part.y * image_h + 0.5))
184 |
185 | pos_move = cnt * back_ground_w / 500
186 | center = (int(center[0] + back_ground_w/1.5 - pos_move/2),
187 | int(center[1] + pos_move))
188 |
189 | centers[i] = center
190 | xs.append(center[0])
191 | ys.append(center[1])
192 | # 绘制关节点
193 | cv.circle(back_ground, centers[i], 3, CocoColors[i], thickness=3, lineType=8, shift=0)
194 |
195 | # 将属于同一人的关节点按照各个部位相连
196 | for pair_order, pair in enumerate(CocoPairsRender):
197 | if pair[0] not in human.body_parts.keys() or pair[1] not in human.body_parts.keys():
198 | continue
199 | cv.line(back_ground, centers[pair[0]], centers[pair[1]], CocoColors[pair_order], 3, cv.LINE_AA)
200 |
201 | @staticmethod
202 | def draw_pose_sequence_horizontally(frame, humans, cnt):
203 | image_h, image_w = frame.shape[:2]
204 | for human in humans:
205 | xs, ys, centers = [], [], {}
206 | # 将所有关节点绘制到图像上
207 | joints_num = CocoPart.Background.value
208 | for i in range(joints_num):
209 | if i not in human.body_parts.keys():
210 | continue
211 |
212 | body_part = human.body_parts[i]
213 | center = (int(body_part.x * image_w + 0.5),
214 | int(body_part.y * image_h + 0.5))
215 |
216 | pos_move = cnt * back_ground_w / 200
217 | center = (int(center[0] + pos_move), center[1])
218 |
219 | centers[i] = center
220 | xs.append(center[0])
221 | ys.append(center[1])
222 | # 绘制关节点
223 | cv.circle(back_ground, centers[i], 3, CocoColors[i], thickness=3, lineType=8, shift=0)
224 |
225 | # 将属于同一人的关节点按照各个部位相连
226 | for pair_order, pair in enumerate(CocoPairsRender):
227 | if pair[0] not in human.body_parts.keys() or pair[1] not in human.body_parts.keys():
228 | continue
229 | cv.line(back_ground, centers[pair[0]], centers[pair[1]], CocoColors[pair_order], 3, cv.LINE_AA)
230 |
231 | def inference(self, npimg):
232 | if npimg is None:
233 | raise Exception('The image does not exist.')
234 |
235 | rois = []
236 | infos = []
237 | # _get_scaled_img
238 | if npimg.shape[:2] != (self.target_size[1], self.target_size[0]):
239 | # resize
240 | npimg = cv.resize(npimg, self.target_size)
241 | rois.extend([npimg])
242 | infos.extend([(0.0, 0.0, 1.0, 1.0)])
243 |
244 | output = self.persistent_sess.run(self.tensor_output, feed_dict={self.tensor_image: rois})
245 |
246 | heat_mats = output[:, :, :, :19]
247 | paf_mats = output[:, :, :, 19:]
248 |
249 | output_h, output_w = output.shape[1:3]
250 | max_ratio_w = max_ratio_h = 10000.0
251 | for info in infos:
252 | max_ratio_w = min(max_ratio_w, info[2])
253 | max_ratio_h = min(max_ratio_h, info[3])
254 | mat_w, mat_h = int(output_w / max_ratio_w), int(output_h / max_ratio_h)
255 |
256 | resized_heat_mat = np.zeros((mat_h, mat_w, 19), dtype=np.float32)
257 | resized_paf_mat = np.zeros((mat_h, mat_w, 38), dtype=np.float32)
258 | resized_cnt_mat = np.zeros((mat_h, mat_w, 1), dtype=np.float32)
259 | resized_cnt_mat += 1e-12
260 |
261 | for heatMat, pafMat, info in zip(heat_mats, paf_mats, infos):
262 | w, h = int(info[2] * mat_w), int(info[3] * mat_h)
263 | heatMat = cv.resize(heatMat, (w, h))
264 | pafMat = cv.resize(pafMat, (w, h))
265 | x, y = int(info[0] * mat_w), int(info[1] * mat_h)
266 | # add up
267 | resized_heat_mat[max(0, y):y + h, max(0, x):x + w, :] = np.maximum(
268 | resized_heat_mat[max(0, y):y + h, max(0, x):x + w, :], heatMat[max(0, -y):, max(0, -x):, :])
269 | resized_paf_mat[max(0, y):y + h, max(0, x):x + w, :] += pafMat[max(0, -y):, max(0, -x):, :]
270 | resized_cnt_mat[max(0, y):y + h, max(0, x):x + w, :] += 1
271 |
272 | self.heatMat = resized_heat_mat
273 | self.pafMat = resized_paf_mat / (np.log(resized_cnt_mat) + 1)
274 |
275 | humans = PoseEstimator.estimate(self.heatMat, self.pafMat)
276 | return humans
277 |
278 |
279 | class PoseEstimator:
280 | heatmap_supress = False
281 | heatmap_gaussian = True
282 | adaptive_threshold = False
283 |
284 | NMS_Threshold = 0.15
285 | Local_PAF_Threshold = 0.2
286 | PAF_Count_Threshold = 5
287 | Part_Count_Threshold = 4
288 | Part_Score_Threshold = 4.5
289 |
290 | PartPair = namedtuple('PartPair', ['score', 'part_idx1', 'part_idx2', 'idx1', 'idx2',
291 | 'coord1', 'coord2', 'score1', 'score2'], verbose=False)
292 |
293 | @staticmethod
294 | def non_max_suppression(plain, window_size=3, threshold=NMS_Threshold):
295 | under_threshold_indices = plain < threshold
296 | plain[under_threshold_indices] = 0
297 | return plain * (plain == maximum_filter(plain, footprint=np.ones((window_size, window_size))))
298 |
299 | @staticmethod
300 | def estimate(heat_mat, paf_mat):
301 | if heat_mat.shape[2] == 19:
302 | heat_mat = np.rollaxis(heat_mat, 2, 0)
303 | if paf_mat.shape[2] == 38:
304 | paf_mat = np.rollaxis(paf_mat, 2, 0)
305 |
306 | if PoseEstimator.heatmap_supress:
307 | heat_mat = heat_mat - heat_mat.min(axis=1).min(axis=1).reshape(19, 1, 1)
308 | heat_mat = heat_mat - heat_mat.min(axis=2).reshape(19, heat_mat.shape[1], 1)
309 |
310 | if PoseEstimator.heatmap_gaussian:
311 | heat_mat = gaussian_filter(heat_mat, sigma=0.5)
312 |
313 | if PoseEstimator.adaptive_threshold:
314 | _NMS_Threshold = max(np.average(heat_mat) * 4.0, PoseEstimator.NMS_Threshold)
315 | _NMS_Threshold = min(_NMS_Threshold, 0.3)
316 | else:
317 | _NMS_Threshold = PoseEstimator.NMS_Threshold
318 |
319 | # extract interesting coordinates using NMS.
320 | coords = [] # [[coords in plane1], [....], ...]
321 | for plain in heat_mat[:-1]:
322 | nms = PoseEstimator.non_max_suppression(plain, 5, _NMS_Threshold)
323 | coords.append(np.where(nms >= _NMS_Threshold))
324 |
325 | # score pairs
326 | pairs_by_conn = list()
327 | for (part_idx1, part_idx2), (paf_x_idx, paf_y_idx) in zip(CocoPairs, CocoPairsNetwork):
328 | pairs = PoseEstimator.score_pairs(
329 | part_idx1, part_idx2,
330 | coords[part_idx1], coords[part_idx2],
331 | paf_mat[paf_x_idx], paf_mat[paf_y_idx],
332 | heatmap=heat_mat,
333 | rescale=(1.0 / heat_mat.shape[2], 1.0 / heat_mat.shape[1])
334 | )
335 |
336 | pairs_by_conn.extend(pairs)
337 |
338 | # merge pairs to human
339 | # pairs_by_conn is sorted by CocoPairs(part importance) and Score between Parts.
340 | humans = [Human([pair]) for pair in pairs_by_conn]
341 | while True:
342 | merge_items = None
343 | for k1, k2 in itertools.combinations(humans, 2):
344 | if k1 == k2:
345 | continue
346 | if k1.is_connected(k2):
347 | merge_items = (k1, k2)
348 | break
349 |
350 | if merge_items is not None:
351 | merge_items[0].merge(merge_items[1])
352 | humans.remove(merge_items[1])
353 | else:
354 | break
355 |
356 | # reject by subset count
357 | humans = [human for human in humans if human.part_count() >= PoseEstimator.PAF_Count_Threshold]
358 | # reject by subset max score
359 | humans = [human for human in humans if human.get_max_score() >= PoseEstimator.Part_Score_Threshold]
360 | return humans
361 |
362 | @staticmethod
363 | def score_pairs(part_idx1, part_idx2, coord_list1, coord_list2, paf_mat_x, paf_mat_y, heatmap, rescale=(1.0, 1.0)):
364 | connection_temp = []
365 |
366 | cnt = 0
367 | for idx1, (y1, x1) in enumerate(zip(coord_list1[0], coord_list1[1])):
368 | for idx2, (y2, x2) in enumerate(zip(coord_list2[0], coord_list2[1])):
369 | score, count = PoseEstimator.get_score(x1, y1, x2, y2, paf_mat_x, paf_mat_y)
370 | cnt += 1
371 | if count < PoseEstimator.PAF_Count_Threshold or score <= 0.0:
372 | continue
373 | connection_temp.append(PoseEstimator.PartPair(
374 | score=score,
375 | part_idx1=part_idx1, part_idx2=part_idx2,
376 | idx1=idx1, idx2=idx2,
377 | coord1=(x1 * rescale[0], y1 * rescale[1]),
378 | coord2=(x2 * rescale[0], y2 * rescale[1]),
379 | score1=heatmap[part_idx1][y1][x1],
380 | score2=heatmap[part_idx2][y2][x2],
381 | ))
382 |
383 | connection = []
384 | used_idx1, used_idx2 = set(), set()
385 | for candidate in sorted(connection_temp, key=lambda x: x.score, reverse=True):
386 | # check not connected
387 | if candidate.idx1 in used_idx1 or candidate.idx2 in used_idx2:
388 | continue
389 | connection.append(candidate)
390 | used_idx1.add(candidate.idx1)
391 | used_idx2.add(candidate.idx2)
392 |
393 | return connection
394 |
395 | @staticmethod
396 | def get_score(x1, y1, x2, y2, paf_mat_x, paf_mat_y):
397 | __num_inter = 10
398 | __num_inter_f = float(__num_inter)
399 | dx, dy = x2 - x1, y2 - y1
400 | normVec = math.sqrt(dx ** 2 + dy ** 2)
401 |
402 | if normVec < 1e-4:
403 | return 0.0, 0
404 |
405 | vx, vy = dx / normVec, dy / normVec
406 |
407 | xs = np.arange(x1, x2, dx / __num_inter_f) if x1 != x2 else np.full((__num_inter,), x1)
408 | ys = np.arange(y1, y2, dy / __num_inter_f) if y1 != y2 else np.full((__num_inter,), y1)
409 | xs = (xs + 0.5).astype(np.int8)
410 | ys = (ys + 0.5).astype(np.int8)
411 |
412 | # without vectorization
413 | pafXs = np.zeros(__num_inter)
414 | pafYs = np.zeros(__num_inter)
415 | for idx, (mx, my) in enumerate(zip(xs, ys)):
416 | pafXs[idx] = paf_mat_x[my][mx]
417 | pafYs[idx] = paf_mat_y[my][mx]
418 |
419 | local_scores = pafXs * vx + pafYs * vy
420 | thidxs = local_scores > PoseEstimator.Local_PAF_Threshold
421 |
422 | return sum(local_scores * thidxs), sum(thidxs)
423 |
424 |
425 | class CocoPart(Enum):
426 | Nose = 0
427 | Neck = 1
428 | RShoulder = 2
429 | RElbow = 3
430 | RWrist = 4
431 | LShoulder = 5
432 | LElbow = 6
433 | LWrist = 7
434 | RHip = 8
435 | RKnee = 9
436 | RAnkle = 10
437 | LHip = 11
438 | LKnee = 12
439 | LAnkle = 13
440 | REye = 14
441 | LEye = 15
442 | REar = 16
443 | LEar = 17
444 | Background = 18
445 |
446 |
447 | FilePath = Path.cwd()
448 | outFilePath = Path(FilePath / "test_out/")
449 | # 有两种输入：height 368 * width 368 （不保持宽高比）；height 368 * width (保持宽高比)
450 | input_width, input_height = 490, 368
451 |
452 | # Usage example: python openpose.py --video=test.mp4
453 | parser = argparse.ArgumentParser(description='OpenPose for pose skeleton in python')
454 | parser.add_argument('--image', help='Path to image file.')
455 | parser.add_argument('--video', help='Path to video file.')
456 | args = parser.parse_args()
457 |
458 | skeleton_estimator = None
459 | nPoints = 18
460 | CocoPairs = [(1, 2), (1, 5), (2, 3), (3, 4), (5, 6), (6, 7), (1, 8),
461 | (8, 9), (9, 10), (1, 11), (11, 12), (12, 13), (1, 0), (0, 14),
462 | (14, 16), (0, 15), (15, 17), (2, 16), (5, 17)] # = 19
463 | CocoPairsRender = CocoPairs[:-2]
464 | CocoPairsNetwork = [(12, 13), (20, 21), (14, 15), (16, 17), (22, 23), (24, 25), (0, 1),
465 | (2, 3), (4, 5), (6, 7), (8, 9), (10, 11), (28, 29), (30, 31),
466 | (34, 35), (32, 33), (36, 37), (18, 19), (26, 27)] # = 19
467 |
468 | # CocoColors = [[0, 100, 255], [0, 100, 255], [0, 255, 255], [0, 100, 255], [0, 255, 255], [0, 100, 255],
469 | # [0, 255, 0], [255, 200, 100], [255, 0, 255], [0, 255, 0], [255, 200, 100], [255, 0, 255],
470 | # [0, 0, 255], [255, 0, 0], [200, 200, 0], [255, 0, 0], [200, 200, 0], [0, 0, 0]]
471 | CocoColors = [[255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0],
472 | [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255],
473 | [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]]
474 |
475 |
476 | def choose_run_mode():
477 | global outFilePath
478 | if args.image:
479 | # Open the image file
480 | if not os.path.isfile(args.image):
481 | print("Input image file ", args.image, " doesn't exist")
482 | sys.exit(1)
483 | cap = cv.VideoCapture(args.image)
484 | outFilePath = str(outFilePath / (args.image[:-4] + '_out.jpg'))
485 | elif args.video:
486 | # Open the video file
487 | if not os.path.isfile(args.video):
488 | print("Input video file ", args.video, " doesn't exist")
489 | sys.exit(1)
490 | cap = cv.VideoCapture(args.video)
491 | outFilePath = str(outFilePath/(args.video[:-4]+'_out.mp4'))
492 | else:
493 | # Webcam input
494 | cap = cv.VideoCapture(0)
495 | outFilePath = str(outFilePath / 'webcam_out.mp4')
496 | return cap
497 |
498 |
499 | def load_pretrain_model():
500 | global skeleton_estimator
501 | skeleton_estimator = TfPoseEstimator(
502 | get_graph_path('VGG_origin'), target_size=(input_width, input_height))
503 |
504 |
505 | def get_graph_path(model_name):
506 | dyn_graph_path = {
507 | 'VGG_origin': str(FilePath/"graph_model_coco/graph_opt.pb"),
508 | 'mobilemet': str(FilePath/"graph_model_coco/graph_opt_mobile.pb")
509 | }
510 | graph_path = dyn_graph_path[model_name]
511 | if not os.path.isfile(graph_path):
512 | raise Exception('Graph file doesn\'t exist, path=%s' % graph_path)
513 | return graph_path
514 |
515 |
516 | if __name__ == "__main__":
517 | cap = choose_run_mode()
518 | load_pretrain_model()
519 | fps = FPS().start()
520 | vid_writer = cv.VideoWriter(outFilePath, cv.VideoWriter_fourcc(*'mp4v'), 15,
521 | (round(cap.get(cv.CAP_PROP_FRAME_WIDTH)),
522 | round(cap.get(cv.CAP_PROP_FRAME_HEIGHT))))
523 |
524 | # 画骨架的序列图所需要的白色背景，可更改背景大小
525 | back_ground_h, back_ground_w = 1500, 2400
526 | back_ground = np.ones((back_ground_h, back_ground_w), dtype=np.uint8)
527 | back_ground = cv.cvtColor(back_ground, cv.COLOR_GRAY2BGR)
528 | back_ground[:, :, :] = 255 # white background
529 | # 间隔interval帧，抽取骨架信息一次
530 | interval = 12
531 |
532 | # 在线FPS计算参数
533 | start_time = time.time()
534 | fps_interval = 1 # 每隔1秒重新计算帧数
535 | fps_count = 0
536 | realtime_fps = 'Starting'
537 |
538 | frame_count = 0
539 | while cv.waitKey(1) < 0:
540 | has_frame, frame = cap.read()
541 | img_copy = np.copy(frame)
542 | if not has_frame:
543 | cv.waitKey(3000)
544 | break
545 |
546 | frame_count += 1
547 | fps_count += 1
548 | humans = skeleton_estimator.inference(frame)
549 | # frame_show = frame
550 | frame_show = TfPoseEstimator.draw_pose_bbox(frame, humans)[0]
551 |
552 | if frame_count % interval == 0 and frame_count <= 180:
553 | TfPoseEstimator.draw_pose_sequence_horizontally(frame, humans, frame_count)
554 |
555 | # FPS的实时显示
556 | fps_show = 'FPS:{0:.4}'.format(realtime_fps)
557 | cv.putText(frame, fps_show, (5, 15), cv.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 2)
558 | if (time.time() - start_time) > fps_interval:
559 | # 计算这个interval过程中的帧数，若interval为1秒，则为FPS
560 | realtime_fps = fps_count / (time.time() - start_time)
561 | fps_count = 0 # 帧数清零
562 | start_time = time.time()
563 |
564 | winName = 'Pose Skeleton from OpenPose'
565 | cv.imshow(winName, frame_show)
566 |
567 | # Write the frame with the detection boxes
568 | if args.image:
569 | cv.imwrite(outFilePath, frame_show)
570 | else:
571 | # vid_writer.write(frame_show)
572 | vid_writer.write(img_copy)
573 |
574 | if not args.image:
575 | fps.stop()
576 | vid_writer.release()
577 | cap.release()
578 | cv.imwrite('skeleton_sequence.jpg', back_ground)
579 | cv.destroyAllWindows()
580 |
--------------------------------------------------------------------------------
/pose-estimator-tensorflow/openpose_tf.py:
--------------------------------------------------------------------------------
1 | import cv2 as cv
2 | import numpy as np
3 | import os
4 | import math
5 | import argparse
6 | import sys
7 | import itertools
8 | import tensorflow as tf
9 | from enum import Enum
10 | from collections import namedtuple
11 | from scipy.ndimage import maximum_filter, gaussian_filter
12 | from imutils.video import FPS
13 | from pathlib import Path
14 |
15 |
16 | class Human:
17 | """
18 | body_parts: list of BodyPart
19 | """
20 | __slots__ = ('body_parts', 'pairs', 'uidx_list')
21 |
22 | def __init__(self, pairs):
23 | self.pairs = []
24 | self.uidx_list = set()
25 | self.body_parts = {}
26 | for pair in pairs:
27 | self.add_pair(pair)
28 |
29 | @staticmethod
30 | def _get_uidx(part_idx, idx):
31 | return '%d-%d' % (part_idx, idx)
32 |
33 | def add_pair(self, pair):
34 | self.pairs.append(pair)
35 | self.body_parts[pair.part_idx1] = BodyPart(Human._get_uidx(pair.part_idx1, pair.idx1),
36 | pair.part_idx1,
37 | pair.coord1[0], pair.coord1[1], pair.score)
38 | self.body_parts[pair.part_idx2] = BodyPart(Human._get_uidx(pair.part_idx2, pair.idx2),
39 | pair.part_idx2,
40 | pair.coord2[0], pair.coord2[1], pair.score)
41 | self.uidx_list.add(Human._get_uidx(pair.part_idx1, pair.idx1))
42 | self.uidx_list.add(Human._get_uidx(pair.part_idx2, pair.idx2))
43 |
44 | def is_connected(self, other):
45 | return len(self.uidx_list & other.uidx_list) > 0
46 |
47 | def merge(self, other):
48 | for pair in other.pairs:
49 | self.add_pair(pair)
50 |
51 | def part_count(self):
52 | return len(self.body_parts.keys())
53 |
54 | def get_max_score(self):
55 | return max([x.score for _, x in self.body_parts.items()])
56 |
57 | def __str__(self):
58 | return ' '.join([str(x) for x in self.body_parts.values()])
59 |
60 |
61 | class BodyPart:
62 | """
63 | part_idx : part index(eg. 0 for nose)
64 | x, y: coordinate of body part
65 | score : confidence score
66 | """
67 | __slots__ = ('uidx', 'part_idx', 'x', 'y', 'score')
68 |
69 | def __init__(self, uidx, part_idx, x, y, score):
70 | self.uidx = uidx
71 | self.part_idx = part_idx
72 | self.x, self.y = x, y
73 | self.score = score
74 |
75 |
76 | class TfPoseEstimator:
77 | ENSEMBLE = 'addup' # average, addup
78 |
79 | def __init__(self, graph_path, target_size=(320, 240)):
80 | self.target_size = target_size
81 |
82 | # load graph
83 | with tf.gfile.GFile(graph_path, 'rb') as f:
84 | graph_def = tf.GraphDef()
85 | graph_def.ParseFromString(f.read())
86 |
87 | self.graph = tf.get_default_graph()
88 | tf.import_graph_def(graph_def, name='TfPoseEstimator')
89 | self.persistent_sess = tf.Session(graph=self.graph)
90 |
91 | self.tensor_image = self.graph.get_tensor_by_name('TfPoseEstimator/image:0')
92 | self.tensor_output = self.graph.get_tensor_by_name('TfPoseEstimator/Openpose/concat_stage7:0')
93 | self.heatMat = self.pafMat = None
94 |
95 | @staticmethod
96 | def draw_pose_bbox(npimg, humans, imgcopy=False):
97 | if imgcopy:
98 | npimg = np.copy(npimg)
99 | image_h, image_w = npimg.shape[:2]
100 | joints, bboxes, xcenter = [], [], []
101 | for human in humans:
102 | xs, ys, centers = [], [], {}
103 | # 将所有关节点绘制到图像上
104 | for i in range(CocoPart.Background.value):
105 | if i not in human.body_parts.keys():
106 | continue
107 |
108 | body_part = human.body_parts[i]
109 | center = (int(body_part.x * image_w + 0.5), int(body_part.y * image_h + 0.5))
110 |
111 | centers[i] = center
112 | xs.append(center[0])
113 | ys.append(center[1])
114 | # 绘制关节点
115 | cv.circle(npimg, center, 3, CocoColors[i], thickness=3, lineType=8, shift=0)
116 | # 将属于同一人的关节点按照各个部位相连
117 | for pair_order, pair in enumerate(CocoPairsRender):
118 | if pair[0] not in human.body_parts.keys() or pair[1] not in human.body_parts.keys():
119 | continue
120 | cv.line(npimg, centers[pair[0]], centers[pair[1]], CocoColors[pair_order], 3, cv.LINE_AA)
121 | # 根据每个人的关节点信息生成ROI区域
122 | xmin = float(min(xs) / image_w)
123 | ymin = float(min(ys) / image_h)
124 | xmax = float(max(xs) / image_w)
125 | ymax = float(max(ys) / image_h)
126 | bboxes.append([xmin, ymin, xmax, ymax, 0.9999])
127 | joints.append(centers)
128 | if 1 in centers:
129 | xcenter.append(centers[1][0])
130 |
131 | # draw bounding_boxes
132 | # x_start, x_end= int(xmin*image_w) - 20, int(xmax*image_w) + 20
133 | # y_start, y_end =int(ymin*image_h) - 15, int(ymax*image_h) + 15
134 | # cv.rectangle(npimg, (x_start, y_start), (x_end, y_end), [0, 250, 0], 3)
135 | return npimg, joints, bboxes, xcenter
136 |
137 | def inference(self, npimg):
138 | if npimg is None:
139 | raise Exception('The image does not exist.')
140 |
141 | rois = []
142 | infos = []
143 | # _get_scaled_img
144 | if npimg.shape[:2] != (self.target_size[1], self.target_size[0]):
145 | # resize
146 | npimg = cv.resize(npimg, self.target_size)
147 | rois.extend([npimg])
148 | infos.extend([(0.0, 0.0, 1.0, 1.0)])
149 |
150 | output = self.persistent_sess.run(self.tensor_output, feed_dict={self.tensor_image: rois})
151 |
152 | heat_mats = output[:, :, :, :19]
153 | paf_mats = output[:, :, :, 19:]
154 |
155 | output_h, output_w = output.shape[1:3]
156 | max_ratio_w = max_ratio_h = 10000.0
157 | for info in infos:
158 | max_ratio_w = min(max_ratio_w, info[2])
159 | max_ratio_h = min(max_ratio_h, info[3])
160 | mat_w, mat_h = int(output_w / max_ratio_w), int(output_h / max_ratio_h)
161 |
162 | resized_heat_mat = np.zeros((mat_h, mat_w, 19), dtype=np.float32)
163 | resized_paf_mat = np.zeros((mat_h, mat_w, 38), dtype=np.float32)
164 | resized_cnt_mat = np.zeros((mat_h, mat_w, 1), dtype=np.float32)
165 | resized_cnt_mat += 1e-12
166 |
167 | for heatMat, pafMat, info in zip(heat_mats, paf_mats, infos):
168 | w, h = int(info[2] * mat_w), int(info[3] * mat_h)
169 | heatMat = cv.resize(heatMat, (w, h))
170 | pafMat = cv.resize(pafMat, (w, h))
171 | x, y = int(info[0] * mat_w), int(info[1] * mat_h)
172 | # add up
173 | resized_heat_mat[max(0, y):y + h, max(0, x):x + w, :] = np.maximum(
174 | resized_heat_mat[max(0, y):y + h, max(0, x):x + w, :], heatMat[max(0, -y):, max(0, -x):, :])
175 | resized_paf_mat[max(0, y):y + h, max(0, x):x + w, :] += pafMat[max(0, -y):, max(0, -x):, :]
176 | resized_cnt_mat[max(0, y):y + h, max(0, x):x + w, :] += 1
177 |
178 | self.heatMat = resized_heat_mat
179 | self.pafMat = resized_paf_mat / (np.log(resized_cnt_mat) + 1)
180 |
181 | humans = PoseEstimator.estimate(self.heatMat, self.pafMat)
182 | return humans
183 |
184 |
185 | class PoseEstimator:
186 | heatmap_supress = False
187 | heatmap_gaussian = True
188 | adaptive_threshold = False
189 |
190 | NMS_Threshold = 0.15
191 | Local_PAF_Threshold = 0.2
192 | PAF_Count_Threshold = 5
193 | Part_Count_Threshold = 4
194 | Part_Score_Threshold = 4.5
195 |
196 | PartPair = namedtuple('PartPair', ['score', 'part_idx1', 'part_idx2', 'idx1', 'idx2',
197 | 'coord1', 'coord2', 'score1', 'score2'], verbose=False)
198 |
199 | @staticmethod
200 | def non_max_suppression(plain, window_size=3, threshold=NMS_Threshold):
201 | under_threshold_indices = plain < threshold
202 | plain[under_threshold_indices] = 0
203 | return plain * (plain == maximum_filter(plain, footprint=np.ones((window_size, window_size))))
204 |
205 | @staticmethod
206 | def estimate(heat_mat, paf_mat):
207 | if heat_mat.shape[2] == 19:
208 | heat_mat = np.rollaxis(heat_mat, 2, 0)
209 | if paf_mat.shape[2] == 38:
210 | paf_mat = np.rollaxis(paf_mat, 2, 0)
211 |
212 | if PoseEstimator.heatmap_supress:
213 | heat_mat = heat_mat - heat_mat.min(axis=1).min(axis=1).reshape(19, 1, 1)
214 | heat_mat = heat_mat - heat_mat.min(axis=2).reshape(19, heat_mat.shape[1], 1)
215 |
216 | if PoseEstimator.heatmap_gaussian:
217 | heat_mat = gaussian_filter(heat_mat, sigma=0.5)
218 |
219 | if PoseEstimator.adaptive_threshold:
220 | _NMS_Threshold = max(np.average(heat_mat) * 4.0, PoseEstimator.NMS_Threshold)
221 | _NMS_Threshold = min(_NMS_Threshold, 0.3)
222 | else:
223 | _NMS_Threshold = PoseEstimator.NMS_Threshold
224 |
225 | # extract interesting coordinates using NMS.
226 | coords = [] # [[coords in plane1], [....], ...]
227 | for plain in heat_mat[:-1]:
228 | nms = PoseEstimator.non_max_suppression(plain, 5, _NMS_Threshold)
229 | coords.append(np.where(nms >= _NMS_Threshold))
230 |
231 | # score pairs
232 | pairs_by_conn = list()
233 | for (part_idx1, part_idx2), (paf_x_idx, paf_y_idx) in zip(CocoPairs, CocoPairsNetwork):
234 | pairs = PoseEstimator.score_pairs(
235 | part_idx1, part_idx2,
236 | coords[part_idx1], coords[part_idx2],
237 | paf_mat[paf_x_idx], paf_mat[paf_y_idx],
238 | heatmap=heat_mat,
239 | rescale=(1.0 / heat_mat.shape[2], 1.0 / heat_mat.shape[1])
240 | )
241 |
242 | pairs_by_conn.extend(pairs)
243 |
244 | # merge pairs to human
245 | # pairs_by_conn is sorted by CocoPairs(part importance) and Score between Parts.
246 | humans = [Human([pair]) for pair in pairs_by_conn]
247 | while True:
248 | merge_items = None
249 | for k1, k2 in itertools.combinations(humans, 2):
250 | if k1 == k2:
251 | continue
252 | if k1.is_connected(k2):
253 | merge_items = (k1, k2)
254 | break
255 |
256 | if merge_items is not None:
257 | merge_items[0].merge(merge_items[1])
258 | humans.remove(merge_items[1])
259 | else:
260 | break
261 |
262 | # reject by subset count
263 | humans = [human for human in humans if human.part_count() >= PoseEstimator.PAF_Count_Threshold]
264 | # reject by subset max score
265 | humans = [human for human in humans if human.get_max_score() >= PoseEstimator.Part_Score_Threshold]
266 | return humans
267 |
268 | @staticmethod
269 | def score_pairs(part_idx1, part_idx2, coord_list1, coord_list2, paf_mat_x, paf_mat_y, heatmap, rescale=(1.0, 1.0)):
270 | connection_temp = []
271 |
272 | cnt = 0
273 | for idx1, (y1, x1) in enumerate(zip(coord_list1[0], coord_list1[1])):
274 | for idx2, (y2, x2) in enumerate(zip(coord_list2[0], coord_list2[1])):
275 | score, count = PoseEstimator.get_score(x1, y1, x2, y2, paf_mat_x, paf_mat_y)
276 | cnt += 1
277 | if count < PoseEstimator.PAF_Count_Threshold or score <= 0.0:
278 | continue
279 | connection_temp.append(PoseEstimator.PartPair(
280 | score=score,
281 | part_idx1=part_idx1, part_idx2=part_idx2,
282 | idx1=idx1, idx2=idx2,
283 | coord1=(x1 * rescale[0], y1 * rescale[1]),
284 | coord2=(x2 * rescale[0], y2 * rescale[1]),
285 | score1=heatmap[part_idx1][y1][x1],
286 | score2=heatmap[part_idx2][y2][x2],
287 | ))
288 |
289 | connection = []
290 | used_idx1, used_idx2 = set(), set()
291 | for candidate in sorted(connection_temp, key=lambda x: x.score, reverse=True):
292 | # check not connected
293 | if candidate.idx1 in used_idx1 or candidate.idx2 in used_idx2:
294 | continue
295 | connection.append(candidate)
296 | used_idx1.add(candidate.idx1)
297 | used_idx2.add(candidate.idx2)
298 |
299 | return connection
300 |
301 | @staticmethod
302 | def get_score(x1, y1, x2, y2, paf_mat_x, paf_mat_y):
303 | __num_inter = 10
304 | __num_inter_f = float(__num_inter)
305 | dx, dy = x2 - x1, y2 - y1
306 | normVec = math.sqrt(dx ** 2 + dy ** 2)
307 |
308 | if normVec < 1e-4:
309 | return 0.0, 0
310 |
311 | vx, vy = dx / normVec, dy / normVec
312 |
313 | xs = np.arange(x1, x2, dx / __num_inter_f) if x1 != x2 else np.full((__num_inter,), x1)
314 | ys = np.arange(y1, y2, dy / __num_inter_f) if y1 != y2 else np.full((__num_inter,), y1)
315 | xs = (xs + 0.5).astype(np.int8)
316 | ys = (ys + 0.5).astype(np.int8)
317 |
318 | # without vectorization
319 | pafXs = np.zeros(__num_inter)
320 | pafYs = np.zeros(__num_inter)
321 | for idx, (mx, my) in enumerate(zip(xs, ys)):
322 | pafXs[idx] = paf_mat_x[my][mx]
323 | pafYs[idx] = paf_mat_y[my][mx]
324 |
325 | local_scores = pafXs * vx + pafYs * vy
326 | thidxs = local_scores > PoseEstimator.Local_PAF_Threshold
327 |
328 | return sum(local_scores * thidxs), sum(thidxs)
329 |
330 |
331 | class CocoPart(Enum):
332 | Nose = 0
333 | Neck = 1
334 | RShoulder = 2
335 | RElbow = 3
336 | RWrist = 4
337 | LShoulder = 5
338 | LElbow = 6
339 | LWrist = 7
340 | RHip = 8
341 | RKnee = 9
342 | RAnkle = 10
343 | LHip = 11
344 | LKnee = 12
345 | LAnkle = 13
346 | REye = 14
347 | LEye = 15
348 | REar = 16
349 | LEar = 17
350 | Background = 18
351 |
352 |
353 | FilePath = Path.cwd()
354 | outFilePath = Path(FilePath / "test_out/")
355 | threshold = 0.1
356 | # winWidth, winHeight = 800, 600
357 | input_width, input_height = 432, 368
358 |
359 | # Usage example: python openpose.py --video=test.mp4
360 | parser = argparse.ArgumentParser(description='OpenPose for pose skeleton in python')
361 | parser.add_argument('--image', help='Path to image file.')
362 | parser.add_argument('--video', help='Path to video file.')
363 | args = parser.parse_args()
364 |
365 | skeleton_estimator = None
366 | nPoints = 18
367 | CocoPairs = [(1, 2), (1, 5), (2, 3), (3, 4), (5, 6), (6, 7), (1, 8), (8, 9), (9, 10), (1, 11),
368 | (11, 12), (12, 13), (1, 0), (0, 14), (14, 16), (0, 15), (15, 17), (2, 16), (5, 17)] # = 19
369 | CocoPairsRender = CocoPairs[:-2]
370 | CocoPairsNetwork = [(12, 13), (20, 21), (14, 15), (16, 17), (22, 23), (24, 25), (0, 1), (2, 3), (4, 5),
371 | (6, 7), (8, 9), (10, 11), (28, 29), (30, 31), (34, 35), (32, 33), (36, 37), (18, 19), (26, 27)] # = 19
372 | CocoColors = [[255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0],
373 | [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255],
374 | [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]]
375 |
376 |
377 | def choose_run_mode():
378 | global outFilePath
379 | if args.image:
380 | # Open the image file
381 | if not os.path.isfile(args.image):
382 | print("Input image file ", args.image, " doesn't exist")
383 | sys.exit(1)
384 | cap = cv.VideoCapture(args.image)
385 | outFilePath = str(outFilePath / (args.image[:-4] + '_out.jpg'))
386 | elif args.video:
387 | # Open the video file
388 | if not os.path.isfile(args.video):
389 | print("Input video file ", args.video, " doesn't exist")
390 | sys.exit(1)
391 | cap = cv.VideoCapture(args.video)
392 | outFilePath = str(outFilePath/(args.video[:-4]+'_out.mp4'))
393 | else:
394 | # Webcam input
395 | cap = cv.VideoCapture(0)
396 | outFilePath = str(outFilePath / 'webcam_out.mp4')
397 | return cap
398 |
399 |
400 | def load_pretrain_model():
401 | global skeleton_estimator
402 | skeleton_estimator = TfPoseEstimator(
403 | get_graph_path('VGG_origin'), target_size=(input_width, input_height))
404 |
405 |
406 | def get_graph_path(model_name):
407 | dyn_graph_path = {
408 | 'VGG_origin': str(FilePath/"graph_model_coco/graph_opt.pb"),
409 | 'mobilemet':str(FilePath/"graph_model_coco/graph_opt_mobile.pb")
410 | }
411 | graph_path = dyn_graph_path[model_name]
412 | if not os.path.isfile(graph_path):
413 | raise Exception('Graph file doesn\'t exist, path=%s' % graph_path)
414 | return graph_path
415 |
416 |
417 | def show_fps(frame):
418 | if not args.image:
419 | fps.update()
420 | fps.stop()
421 | fps_label = "FPS: {:.2f}".format(fps.fps())
422 | cv.putText(frame, fps_label, (5, 15), cv.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 2)
423 |
424 |
425 | if __name__ == "__main__":
426 | cap = choose_run_mode()
427 | load_pretrain_model()
428 | fps = FPS().start()
429 | vid_writer = cv.VideoWriter(outFilePath, cv.VideoWriter_fourcc(*'mp4v'), 30,
430 | (round(cap.get(cv.CAP_PROP_FRAME_WIDTH)),
431 | round(cap.get(cv.CAP_PROP_FRAME_HEIGHT))))
432 |
433 | while cv.waitKey(1) < 0:
434 | has_frame, frame = cap.read()
435 | # frame = cv.resize(frame, (winWidth, winHeight))
436 | if not has_frame:
437 | cv.waitKey(3000)
438 | break
439 | humans = skeleton_estimator.inference(frame)
440 | frame_show = TfPoseEstimator.draw_pose_bbox(frame, humans)[0]
441 | # 显示实时FPS
442 | show_fps(frame_show)
443 | winName = 'Pose Skeleton from OpenPose'
444 | # cv.namedWindow(winName, cv.WINDOW_NORMAL)
445 | cv.imshow(winName, frame_show)
446 |
447 | # Write the frame with the detection boxes
448 | if args.image:
449 | cv.imwrite(outFilePath, frame_show)
450 | else:
451 | vid_writer.write(frame_show)
452 |
453 | if not args.image:
454 | vid_writer.release()
455 | cap.release()
456 | cv.destroyAllWindows()
457 |
--------------------------------------------------------------------------------
/pose-estimator-tensorflow/test_out/README.md:
--------------------------------------------------------------------------------
1 | This directory is for saving the after-processing outputs.
2 |
--------------------------------------------------------------------------------
/pose-estimator-tensorflow/test_out/skeleton_sequence.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/pose-estimator-tensorflow/test_out/skeleton_sequence.jpg
--------------------------------------------------------------------------------
/pose-estimator-tensorflow/test_out/skeleton_sequence_horizontal.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/pose-estimator-tensorflow/test_out/skeleton_sequence_horizontal.jpg
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/README.md:
--------------------------------------------------------------------------------
1 |
2 | ### Usage Examples :
3 | Put the test file (iamge or video) under the same directory
4 |
5 | - `python3 estimator_multi_people.py --image=test.jpg`
6 | - `python3 estimator_multi_people.py --video=test.mp4`
7 | - if no argument provided, it starts the webcam.
8 |
9 | ### Note
10 | You can refer [here](https://www.learnopencv.com/multi-person-pose-estimation-in-opencv-using-openpose/) in english, or [here](https://blog.csdn.net/qq_27158179/article/details/82717821) in chinese.
11 |
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/estimator_multi_people.py:
--------------------------------------------------------------------------------
1 | import os
2 | import sys
3 | import cv2 as cv
4 | import numpy as np
5 | from pathlib import Path
6 | import argparse
7 | from imutils.video import FPS
8 |
9 |
10 | # reference: https://blog.csdn.net/qq_27158179/article/details/82717821
11 |
12 | # 相关参数初始化
13 | FilePath = Path.cwd()
14 | out_file_path = Path(FilePath / "test_out/")
15 |
16 | joints_list_with_id = [] # 储存frame中所有joints, [[（x, y, conf, id), ...], [（x, y, conf, id), ...]]
17 | joints_list = np.zeros((0, 3)) # 储存frame中所有joints, [[x, y, conf], ...]
18 | joint_id = 0 # frame中所有joints的id
19 |
20 | # 以COCO骨架格式为例
21 | proto_file = str(FilePath / "model/coco/pose_deploy_linevec_faster_4_stages.prototxt")
22 | weights_file = str(FilePath / "model/coco/pose_iter_440000.caffemodel")
23 |
24 | coco_num = 18
25 | joints_mapping = ['Nose', 'Neck', 'R-Sho', 'R-Elb', 'R-Wr', 'L-Sho', 'L-Elb', 'L-Wr',
26 | 'R-Hip', 'R-Knee', 'R-Ank', 'L-Hip', 'L-Knee', 'L-Ank', 'R-Eye', 'L-Eye',
27 | 'R-Ear', 'L-Ear', 'Background']
28 | POSE_PAIRS = [[1, 2], [1, 5], [2, 3], [3, 4], [5, 6], [6, 7],
29 | [1, 8], [8, 9], [9, 10], [1, 11], [11, 12], [12, 13],
30 | [1, 0], [0, 14], [14, 16], [0, 15], [15, 17], [2, 17], [5, 16]]
31 | # index of PAFs corresponding to the POSE_PAIRS
32 | # e.g for POSE_PAIR(1,2), the PAFs are located at indices (31,32) of output, Similarly, (1,5) -> (39,40) and so on.
33 | MAP_INDEX = [[31, 32], [39, 40], [33, 34], [35, 36], [41, 42], [43, 44],
34 | [19, 20], [21, 22], [23, 24], [25, 26], [27, 28], [29, 30],
35 | [47, 48], [49, 50], [53, 54], [51, 52], [55, 56], [37, 38], [45, 46]]
36 |
37 | COLORS = [[0, 100, 255], [0, 100, 255], [0, 255, 255], [0, 100, 255], [0, 255, 255], [0, 100, 255],
38 | [0, 255, 0], [255, 200, 100], [255, 0, 255], [0, 255, 0], [255, 200, 100], [255, 0, 255],
39 | [0, 0, 255], [255, 0, 0], [200, 200, 0], [255, 0, 0], [200, 200, 0], [0, 0, 0]]
40 |
41 |
42 | def choose_run_mode():
43 | """
44 | 选择输入：image/video/webcam
45 | """
46 | global out_file_path
47 | if args.image:
48 | # Open the image file
49 | if not os.path.isfile(args.image):
50 | print("Input image file ", args.image, " doesn't exist")
51 | sys.exit(1)
52 | cap = cv.VideoCapture(args.image)
53 | out_file_path = str(out_file_path / (args.image[:-4] + '_out.jpg'))
54 | elif args.video:
55 | # Open the video file
56 | if not os.path.isfile(args.video):
57 | print("Input video file ", args.video, " doesn't exist")
58 | sys.exit(1)
59 | cap = cv.VideoCapture(args.video)
60 | out_file_path = str(out_file_path/(args.video[:-4]+'_out.mp4'))
61 | else:
62 | # Webcam input
63 | cap = cv.VideoCapture(0)
64 | out_file_path = str(out_file_path / 'webcam_out.mp4')
65 | return cap
66 |
67 |
68 | def get_joints(prob_map, threshold=0.15):
69 | """
70 | :param prob_map: image中某一类joint的map
71 | :param threshold: 以这个threshold过滤掉map中置信度较低的部分
72 | :return: 所有该joint类别的集合，形式为(x, y, conf)
73 | """
74 | # 以较低的threshold值，提取出可能为某关节点的所有像素区域
75 | map_smooth = cv.GaussianBlur(prob_map, (3, 3), 0, 0)
76 | # 生成这个区域的mask
77 | map_mask = np.uint8(map_smooth > threshold)
78 | joints = []
79 | # 围绕mask画轮廓contour
80 | _, contours, _ = cv.findContours(map_mask, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)
81 | # 针对每一个person某joint的contour，找到confidence最大的一个点，作为估计的关节点位置
82 | for cnt in contours:
83 | blob_mask = np.zeros(map_mask.shape)
84 | blob_mask = cv.fillConvexPoly(blob_mask, cnt, 1)
85 | masked_prob_map = map_smooth * blob_mask
86 | _, max_val, _, max_loc = cv.minMaxLoc(masked_prob_map)
87 | joints.append(max_loc + (prob_map[max_loc[1], max_loc[0]],))
88 | # 返回
89 | return joints
90 |
91 |
92 | def get_valid_pairs(output):
93 | """
94 | Find valid / invalid connections between the different joints of a all persons present
95 | """
96 | valid_pairs = []
97 | invalid_pairs = []
98 | n_interp_samples = 10 # 插值点数目
99 | paf_threshold = 0.2
100 | conf_threshold = 0.5
101 | # loop for every POSE_PAIR
102 | for k in range(len(MAP_INDEX)):
103 | # a->b constitute a limb
104 | paf_a = output[0, MAP_INDEX[k][0], :, :]
105 | # print(paf_a.shape)
106 | paf_b = output[0, MAP_INDEX[k][1], :, :]
107 | paf_a = cv.resize(paf_a, (frameWidth, frameHeight))
108 | paf_b = cv.resize(paf_b, (frameWidth, frameHeight))
109 |
110 | # Find the joints for the first and second limb
111 | # cand_a为某一joint的列表， cand_b为另一与之相连接的joint的列表
112 | cand_a = joints_list_with_id[POSE_PAIRS[k][0]]
113 | cand_b = joints_list_with_id[POSE_PAIRS[k][1]]
114 | # 在完美检测到frame中所有joints的情况下， n_a = n_b = len(persons)
115 | n_a = len(cand_a)
116 | n_b = len(cand_b)
117 |
118 | # If joints for the joint-pair is detected
119 | # check every joint in cand_a with every joint in cand_b
120 | if n_a != 0 and n_b != 0:
121 | valid_pair = np.zeros((0, 3))
122 | for i in range(n_a):
123 | max_j = -1
124 | max_score = -1
125 | found = False
126 | for j in range(n_b):
127 | # Calculate the distance vector between the two joints
128 | distance_ij = np.subtract(cand_b[j][:2], cand_a[i][:2])
129 | # 求二范数，即求模，算两点距离
130 | norm = np.linalg.norm(distance_ij)
131 | if norm:
132 | # 距离不为零的话， 缩放到单位向量
133 | distance_ij = distance_ij / norm
134 | else:
135 | continue
136 |
137 | # Find p(u)，在连接两joints的直线上创建一个n_interp_samples插值点的数组
138 | interp_coord = list(zip(np.linspace(cand_a[i][0], cand_b[j][0], num=n_interp_samples),
139 | np.linspace(cand_a[i][1], cand_b[j][1], num=n_interp_samples)))
140 | # Find the PAF values at a set of interpolated points between the joints
141 | paf_interp = []
142 | for m in range(len(interp_coord)):
143 | paf_interp.append([paf_a[int(round(interp_coord[m][1])), int(round(interp_coord[m][0]))],
144 | paf_b[int(round(interp_coord[m][1])), int(round(interp_coord[m][0]))]])
145 | # Find E
146 | paf_scores = np.dot(paf_interp, distance_ij)
147 | avg_paf_score = sum(paf_scores)/len(paf_scores)
148 |
149 | # Check if the connection is valid
150 | # If the fraction of interpolated vectors aligned with PAF is higher then threshold -> Valid Pair
151 | if (len(np.where(paf_scores > paf_threshold)[0]) / n_interp_samples) > conf_threshold:
152 | if avg_paf_score > max_score:
153 | # 如果这些点中有70%大于conf threshold，则把这一对当成有效
154 | max_j = j
155 | max_score = avg_paf_score
156 | found = True
157 | # Append the connection to the list
158 | if found:
159 | valid_pair = np.append(valid_pair, [[cand_a[i][3], cand_b[max_j][3], max_score]], axis=0)
160 |
161 | # Append the detected connections to the global list
162 | valid_pairs.append(valid_pair)
163 | # If no joints are detected
164 | else:
165 | # print("No Connection : k = {}".format(k))
166 | invalid_pairs.append(k)
167 | valid_pairs.append([])
168 | return valid_pairs, invalid_pairs
169 |
170 |
171 | def get_personwise_joints(valid_pairs, invalid_pairs):
172 | """
173 | This function creates a list of joints belonging to each person
174 | For each detected valid pair, it assigns the joint(s) to a person
175 | """
176 | # the last number in each row is the overall score
177 | personwise_joints = -1 * np.ones((0, 19))
178 | # print(personwise_joints.shape)
179 |
180 | for k in range(len(MAP_INDEX)):
181 | if k not in invalid_pairs:
182 | part_as = valid_pairs[k][:, 0]
183 | part_bs = valid_pairs[k][:, 1]
184 | index_a, index_b = np.array(POSE_PAIRS[k])
185 |
186 | for i in range(len(valid_pairs[k])):
187 | found = False
188 | person_idx = -1
189 | for j in range(len(personwise_joints)):
190 | if personwise_joints[j][index_a] == part_as[i]:
191 | person_idx = j
192 | found = True
193 | break
194 |
195 | if found:
196 | personwise_joints[person_idx][index_b] = part_bs[i]
197 | personwise_joints[person_idx][-1] += joints_list[part_bs[i].astype(int), 2] + valid_pairs[k][i][2]
198 |
199 | # if find no partA in the subset, create a new subset
200 | elif not found and k < 17:
201 | row = -1 * np.ones(19)
202 | row[index_a] = part_as[i]
203 | row[index_b] = part_bs[i]
204 | # add the joint_scores for the two joints and the paf_score
205 | row[-1] = sum(joints_list[valid_pairs[k][i, :2].astype(int), 2]) + valid_pairs[k][i][2]
206 | personwise_joints = np.vstack([personwise_joints, row])
207 | return personwise_joints
208 |
209 |
210 | def load_pretrained_model():
211 | """
212 | 加载预训练后的POSE的caffe模型
213 | """
214 | net = cv.dnn.readNetFromCaffe(proto_file, weights_file)
215 | # 调用GPU模块，但目前opencv仅支持intel GPU
216 | # tested with Intel GPUs only, or it will automatically switch to CPU
217 | net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
218 | net.setPreferableTarget(cv.dnn.DNN_TARGET_OPENCL)
219 | return net
220 |
221 |
222 | def show_joints(frame):
223 | for i in range(coco_num):
224 | for j in range(len(joints_list_with_id[i])):
225 | cv.circle(frame, joints_list_with_id[i][j][0:2], 5, COLORS[i], -1, cv.LINE_AA)
226 | cv.imshow("joints", frame)
227 |
228 |
229 | if __name__ == '__main__':
230 | # Usage example: python estimator_single_person.py --video=test.mp4
231 | parser = argparse.ArgumentParser(description='OpenPose for pose skeleton in python')
232 | parser.add_argument('--image', help='Path to image file.')
233 | parser.add_argument('--video', help='Path to video file.')
234 | args = parser.parse_args()
235 |
236 | cap = choose_run_mode()
237 | net = load_pretrained_model()
238 | fps = FPS().start()
239 | vid_writer = cv.VideoWriter(out_file_path, cv.VideoWriter_fourcc(*'mp4v'), 30,
240 | (round(cap.get(cv.CAP_PROP_FRAME_WIDTH)),
241 | round(cap.get(cv.CAP_PROP_FRAME_HEIGHT))))
242 |
243 | while cv.waitKey(1) < 0:
244 | hasFrame, frame = cap.read()
245 | if not hasFrame:
246 | print("Output file is stored as ", out_file_path)
247 | cv.waitKey(3000)
248 | break
249 |
250 | frameHeight, frameWidth = frame.shape[:2]
251 |
252 | # Fix the input Height and get the width according to the Aspect Ratio
253 | # 有两种输入：height 368 * width 368 （不保持宽高比）；height 368 * width (保持宽高比)
254 | inHeight, inwidth = 368, 368
255 | inWidth = int((inHeight/frameHeight)*frameWidth) # 源代码中是选择 保持宽高比
256 |
257 | # 首先，我们将像素值标准化为（0,1）。然后我们指定图像的尺寸。接下来，要减去的平均值，即（0,0,0）
258 | # 不同算法及训练模型的blobFromImage参数不同，可访问opencv的github地址查询
259 | # https://github.com/opencv/opencv/tree/master/samples/dnn
260 | inpBlob = cv.dnn.blobFromImage(frame, 1.0 / 255, (inWidth, inHeight), (0, 0, 0), swapRB=False, crop=False)
261 | net.setInput(inpBlob)
262 | # output的[0, i, :, :,]的前19个矩阵为19个（包含一个background）关节点的置信map
263 | # 后38个矩阵为每两个相连接关节点之间形成的PAF矩阵
264 | output = net.forward()
265 |
266 | # 找到image中所有可能的关节点joint的坐标
267 | for part in range(coco_num):
268 | # 按关节点的序列分别寻找
269 | joint_prob_map = output[0, part, :, :]
270 | joint_prob_map = cv.resize(joint_prob_map, (frame.shape[1], frame.shape[0]))
271 | # 调用get_joints函数，获取image中的所有人的某一joint，len(joints)==len(persons)
272 | joints = get_joints(joint_prob_map)
273 |
274 | joints_with_id = []
275 | for i in range(len(joints)):
276 | # joints集合，竖直叠加（无id）
277 | joints_list = np.vstack([joints_list, joints[i]])
278 | # joints集合（有id）
279 | joints_with_id.append(joints[i] + (joint_id,))
280 | joint_id += 1
281 | joints_list_with_id.append(joints_with_id)
282 |
283 | # # 标出所有joints位置(no joints connection)
284 | # show_joints(frame.copy())
285 |
286 | # 画多人skeleton（joints connection）
287 | valid_pairs, invalid_pairs = get_valid_pairs(output)
288 | personwise_joints = get_personwise_joints(valid_pairs, invalid_pairs)
289 |
290 | for i in range(17):
291 | for n in range(len(personwise_joints)):
292 | index = personwise_joints[n][np.array(POSE_PAIRS[i])]
293 | if -1 in index:
294 | continue
295 | B = np.int32(joints_list[index.astype(int), 0])
296 | A = np.int32(joints_list[index.astype(int), 1])
297 | cv.line(frame, (B[0], A[0]), (B[1], A[1]), COLORS[i], 3, cv.LINE_AA)
298 |
299 | # Write the frame with the detection boxes
300 | if args.image:
301 | cv.imwrite(out_file_path, frame)
302 | else:
303 | vid_writer.write(frame)
304 |
305 | cv.imshow("Detected Pose", frame)
306 |
307 | if not args.image:
308 | vid_writer.release()
309 | cap.release()
310 | cv.destroyAllWindows()
311 |
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/estimator_single_person.py:
--------------------------------------------------------------------------------
1 | import cv2 as cv
2 | import os
3 | import argparse
4 | import sys
5 | from imutils.video import FPS
6 | from pathlib import Path
7 |
8 | # 相关参数初始化
9 | FilePath = Path.cwd()
10 | out_file_path = Path(FilePath / "test_out/")
11 | threshold = 0.1
12 | input_width, input_height = 368, 368
13 | lineWidthMultiper = 2
14 |
15 | proto_file = ""
16 | weights_file = ""
17 | nPoints = 0
18 | POSE_PAIRS = []
19 |
20 |
21 | def choose_pose(model='COCO'):
22 | """
23 | 选择输出的骨架模型, 可选：MPI / COCO / BODY_25
24 | """
25 | global proto_file, weights_file, nPoints, POSE_PAIRS
26 | if model is "COCO":
27 | proto_file = str(FilePath / "model/coco/pose_deploy_linevec_faster_4_stages.prototxt")
28 | weights_file = str(FilePath / "model/coco/pose_iter_440000.caffemodel")
29 | nPoints = 18
30 | POSE_PAIRS = [[1, 0], [1, 2], [1, 5], [2, 3], [3, 4], [5, 6], [6, 7], [1, 8], [8, 9], [9, 10], [1, 11], [11, 12],
31 | [12, 13], [0, 14], [0, 15], [14, 16], [15, 17]]
32 | elif model is "MPI":
33 | proto_file = str(FilePath / "model/mpi/pose_deploy_linevec_faster_4_stages.prototxt")
34 | weights_file = str(FilePath / "model/mpi/pose_iter_440000.caffemodel")
35 | nPoints = 15
36 | POSE_PAIRS = [[0, 1], [1, 2], [2, 3], [3, 4], [1, 5], [5, 6], [6, 7], [1, 14], [14, 8], [8, 9], [9, 10], [14, 11],
37 | [11, 12], [12, 13]]
38 | elif model is "BODY_25":
39 | proto_file = str(FilePath / "model/body_25/pose_deploy.prototxt")
40 | weights_file = str(FilePath / "model/body_25/pose_iter_584000.caffemodel")
41 | nPoints = 25
42 | POSE_PAIRS = [[1, 0], [1, 2], [1, 5], [2, 3], [3, 4], [5, 6], [6, 7], [1, 8], [8, 9], [9, 10], [10, 11], [11, 24],
43 | [11, 22], [22, 23], [8, 12], [12, 13], [13, 14], [14, 21], [14, 19], [19, 20], [0, 15], [15, 17],
44 | [0, 16], [16, 18]]
45 |
46 |
47 | def choose_run_mode():
48 | """
49 | 选择输入：image/video/webcam
50 | """
51 | global out_file_path
52 | if args.image:
53 | # Open the image file
54 | if not os.path.isfile(args.image):
55 | print("Input image file ", args.image, " doesn't exist")
56 | sys.exit(1)
57 | cap = cv.VideoCapture(args.image)
58 | out_file_path = str(out_file_path / (args.image[:-4] + '_out.jpg'))
59 | elif args.video:
60 | # Open the video file
61 | if not os.path.isfile(args.video):
62 | print("Input video file ", args.video, " doesn't exist")
63 | sys.exit(1)
64 | cap = cv.VideoCapture(args.video)
65 | out_file_path = str(out_file_path/(args.video[:-4]+'_out.mp4'))
66 | else:
67 | # Webcam input
68 | cap = cv.VideoCapture(0)
69 | out_file_path = str(out_file_path / 'webcam_out.mp4')
70 | return cap
71 |
72 |
73 | def load_pretrain_model():
74 | """
75 | 加载预训练后的POSE的caffe模型
76 | """
77 | net = cv.dnn.readNetFromCaffe(proto_file, weights_file)
78 | print('POSE caffe model loaded successfully')
79 | # 调用GPU模块，但目前opencv仅支持intel GPU
80 | # tested with Intel GPUs only, or it will automatically switch to CPU
81 | net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
82 | net.setPreferableTarget(cv.dnn.DNN_TARGET_OPENCL)
83 | return net
84 |
85 |
86 | def show_fps():
87 | if not args.image:
88 | fps.update()
89 | fps.stop()
90 | fps_label = "FPS: {:.2f}".format(fps.fps())
91 | cv.putText(frame, fps_label, (0, origin_h - 25), cv.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), lineWidthMultiper)
92 |
93 |
94 | if __name__ == "__main__":
95 | # Usage example: python estimator_single_person.py --video=test.mp4
96 | parser = argparse.ArgumentParser(description='OpenPose for pose skeleton in python')
97 | parser.add_argument('--image', help='Path to image file.')
98 | parser.add_argument('--video', help='Path to video file.')
99 | args = parser.parse_args()
100 |
101 | choose_pose('COCO')
102 | cap = choose_run_mode()
103 | net = load_pretrain_model()
104 | fps = FPS().start()
105 | vid_writer = cv.VideoWriter(out_file_path, cv.VideoWriter_fourcc(*'mp4v'), 30,
106 | (round(cap.get(cv.CAP_PROP_FRAME_WIDTH)),
107 | round(cap.get(cv.CAP_PROP_FRAME_HEIGHT))))
108 | while cv.waitKey(1) < 0:
109 | hasFrame, frame = cap.read()
110 | if not hasFrame:
111 | print("Output file is stored as ", out_file_path)
112 | cv.waitKey(3000)
113 | break
114 |
115 | origin_h, origin_w = frame.shape[:2]
116 | # 首先，我们将像素值标准化为（0,1）。然后我们指定图像的尺寸。接下来，要减去的平均值，即（0,0,0）
117 | # 不同算法及训练模型的blobFromImage参数不同，可访问opencv的github地址查询
118 | # https://github.com/opencv/opencv/tree/master/samples/dnn
119 | blob = cv.dnn.blobFromImage(frame, 1.0 / 255, (input_width, input_height), 0, swapRB=False, crop=False)
120 | net.setInput(blob)
121 | # 第一个维度是图像ID（如果将多个图像传递给网络）。
122 | # 第二个维度表示关节点的索引。该模型生成Confidence Maps和Part Affinity Maps。
123 | # 对于COCO模型，它由57个部分组成: 18个关键点置信度map + 1个背景 + 19 * 2部分亲和力图。
124 | # 第三个维度是输出映射map的高度。
125 | # 第四个维度是输出映射map的宽度。
126 | detections = net.forward()
127 | H = detections.shape[2]
128 | W = detections.shape[3]
129 | # 存储关节点
130 | points = []
131 |
132 | for i in range(nPoints):
133 | probility_map = detections[0, i, :, :]
134 | #
135 | min_value, confidence, min_loc, point = cv.minMaxLoc(probility_map)
136 | #
137 | x = int(origin_w * (point[0] / W))
138 | y = int(origin_h * (point[1] / H))
139 |
140 | if confidence > threshold:
141 | cv.circle(frame, (x, y), lineWidthMultiper*3, (0, 255, 255), -1, cv.FILLED)
142 | # cv.putText(frame, "{}".format(i), (x, y-15), cv.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 255), 1, cv.LINE_AA)
143 | points.append((x, y))
144 | else:
145 | points.append(None)
146 |
147 | # 画骨架
148 | for pair in POSE_PAIRS:
149 | A, B = pair[0], pair[1]
150 | if points[A] and points[B]:
151 | cv.line(frame, points[A], points[B], (0, 0, 255), lineWidthMultiper, cv.LINE_AA)
152 | # cv.circle(frame, points[A], 8, (0, 0, 255), thickness=-1, lineType=cv.FILLED)
153 | # cv.circle(frame, points[B], 8, (0, 0, 255), thickness=-1, lineType=cv.FILLED)
154 | # 显示实时FPS
155 | show_fps()
156 |
157 | # Write the frame with the detection boxes
158 | if args.image:
159 | cv.imwrite(out_file_path, frame)
160 | else:
161 | vid_writer.write(frame)
162 |
163 | winName = 'Pose Skeleton from OpenPose'
164 | # cv.namedWindow(winName, cv.WINDOW_NORMAL)
165 | cv.imshow(winName, frame)
166 |
167 | if not args.image:
168 | vid_writer.release()
169 | cap.release()
170 | cv.destroyAllWindows()
171 |
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/model/body_25/keypoints_pose_25.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/pose-estimator-using-caffemodel/model/body_25/keypoints_pose_25.png
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/model/coco/keypoints_pose_18.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/pose-estimator-using-caffemodel/model/coco/keypoints_pose_18.png
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/model/coco/pose_deploy_linevec_faster_4_stages.prototxt:
--------------------------------------------------------------------------------
1 | input: "image"
2 | input_dim: 1
3 | input_dim: 3
4 | input_dim: 1 # This value will be defined at runtime
5 | input_dim: 1 # This value will be defined at runtime
6 | layer {
7 | name: "conv1_1"
8 | type: "Convolution"
9 | bottom: "image"
10 | top: "conv1_1"
11 | param {
12 | lr_mult: 1.0
13 | decay_mult: 1
14 | }
15 | param {
16 | lr_mult: 2.0
17 | decay_mult: 0
18 | }
19 | convolution_param {
20 | num_output: 64
21 | pad: 1
22 | kernel_size: 3
23 | weight_filler {
24 | type: "gaussian"
25 | std: 0.01
26 | }
27 | bias_filler {
28 | type: "constant"
29 | }
30 | }
31 | }
32 | layer {
33 | name: "relu1_1"
34 | type: "ReLU"
35 | bottom: "conv1_1"
36 | top: "conv1_1"
37 | }
38 | layer {
39 | name: "conv1_2"
40 | type: "Convolution"
41 | bottom: "conv1_1"
42 | top: "conv1_2"
43 | param {
44 | lr_mult: 1.0
45 | decay_mult: 1
46 | }
47 | param {
48 | lr_mult: 2.0
49 | decay_mult: 0
50 | }
51 | convolution_param {
52 | num_output: 64
53 | pad: 1
54 | kernel_size: 3
55 | weight_filler {
56 | type: "gaussian"
57 | std: 0.01
58 | }
59 | bias_filler {
60 | type: "constant"
61 | }
62 | }
63 | }
64 | layer {
65 | name: "relu1_2"
66 | type: "ReLU"
67 | bottom: "conv1_2"
68 | top: "conv1_2"
69 | }
70 | layer {
71 | name: "pool1_stage1"
72 | type: "Pooling"
73 | bottom: "conv1_2"
74 | top: "pool1_stage1"
75 | pooling_param {
76 | pool: MAX
77 | kernel_size: 2
78 | stride: 2
79 | }
80 | }
81 | layer {
82 | name: "conv2_1"
83 | type: "Convolution"
84 | bottom: "pool1_stage1"
85 | top: "conv2_1"
86 | param {
87 | lr_mult: 1.0
88 | decay_mult: 1
89 | }
90 | param {
91 | lr_mult: 2.0
92 | decay_mult: 0
93 | }
94 | convolution_param {
95 | num_output: 128
96 | pad: 1
97 | kernel_size: 3
98 | weight_filler {
99 | type: "gaussian"
100 | std: 0.01
101 | }
102 | bias_filler {
103 | type: "constant"
104 | }
105 | }
106 | }
107 | layer {
108 | name: "relu2_1"
109 | type: "ReLU"
110 | bottom: "conv2_1"
111 | top: "conv2_1"
112 | }
113 | layer {
114 | name: "conv2_2"
115 | type: "Convolution"
116 | bottom: "conv2_1"
117 | top: "conv2_2"
118 | param {
119 | lr_mult: 1.0
120 | decay_mult: 1
121 | }
122 | param {
123 | lr_mult: 2.0
124 | decay_mult: 0
125 | }
126 | convolution_param {
127 | num_output: 128
128 | pad: 1
129 | kernel_size: 3
130 | weight_filler {
131 | type: "gaussian"
132 | std: 0.01
133 | }
134 | bias_filler {
135 | type: "constant"
136 | }
137 | }
138 | }
139 | layer {
140 | name: "relu2_2"
141 | type: "ReLU"
142 | bottom: "conv2_2"
143 | top: "conv2_2"
144 | }
145 | layer {
146 | name: "pool2_stage1"
147 | type: "Pooling"
148 | bottom: "conv2_2"
149 | top: "pool2_stage1"
150 | pooling_param {
151 | pool: MAX
152 | kernel_size: 2
153 | stride: 2
154 | }
155 | }
156 | layer {
157 | name: "conv3_1"
158 | type: "Convolution"
159 | bottom: "pool2_stage1"
160 | top: "conv3_1"
161 | param {
162 | lr_mult: 1.0
163 | decay_mult: 1
164 | }
165 | param {
166 | lr_mult: 2.0
167 | decay_mult: 0
168 | }
169 | convolution_param {
170 | num_output: 256
171 | pad: 1
172 | kernel_size: 3
173 | weight_filler {
174 | type: "gaussian"
175 | std: 0.01
176 | }
177 | bias_filler {
178 | type: "constant"
179 | }
180 | }
181 | }
182 | layer {
183 | name: "relu3_1"
184 | type: "ReLU"
185 | bottom: "conv3_1"
186 | top: "conv3_1"
187 | }
188 | layer {
189 | name: "conv3_2"
190 | type: "Convolution"
191 | bottom: "conv3_1"
192 | top: "conv3_2"
193 | param {
194 | lr_mult: 1.0
195 | decay_mult: 1
196 | }
197 | param {
198 | lr_mult: 2.0
199 | decay_mult: 0
200 | }
201 | convolution_param {
202 | num_output: 256
203 | pad: 1
204 | kernel_size: 3
205 | weight_filler {
206 | type: "gaussian"
207 | std: 0.01
208 | }
209 | bias_filler {
210 | type: "constant"
211 | }
212 | }
213 | }
214 | layer {
215 | name: "relu3_2"
216 | type: "ReLU"
217 | bottom: "conv3_2"
218 | top: "conv3_2"
219 | }
220 | layer {
221 | name: "conv3_3"
222 | type: "Convolution"
223 | bottom: "conv3_2"
224 | top: "conv3_3"
225 | param {
226 | lr_mult: 1.0
227 | decay_mult: 1
228 | }
229 | param {
230 | lr_mult: 2.0
231 | decay_mult: 0
232 | }
233 | convolution_param {
234 | num_output: 256
235 | pad: 1
236 | kernel_size: 3
237 | weight_filler {
238 | type: "gaussian"
239 | std: 0.01
240 | }
241 | bias_filler {
242 | type: "constant"
243 | }
244 | }
245 | }
246 | layer {
247 | name: "relu3_3"
248 | type: "ReLU"
249 | bottom: "conv3_3"
250 | top: "conv3_3"
251 | }
252 | layer {
253 | name: "conv3_4"
254 | type: "Convolution"
255 | bottom: "conv3_3"
256 | top: "conv3_4"
257 | param {
258 | lr_mult: 1.0
259 | decay_mult: 1
260 | }
261 | param {
262 | lr_mult: 2.0
263 | decay_mult: 0
264 | }
265 | convolution_param {
266 | num_output: 256
267 | pad: 1
268 | kernel_size: 3
269 | weight_filler {
270 | type: "gaussian"
271 | std: 0.01
272 | }
273 | bias_filler {
274 | type: "constant"
275 | }
276 | }
277 | }
278 | layer {
279 | name: "relu3_4"
280 | type: "ReLU"
281 | bottom: "conv3_4"
282 | top: "conv3_4"
283 | }
284 | layer {
285 | name: "pool3_stage1"
286 | type: "Pooling"
287 | bottom: "conv3_4"
288 | top: "pool3_stage1"
289 | pooling_param {
290 | pool: MAX
291 | kernel_size: 2
292 | stride: 2
293 | }
294 | }
295 | layer {
296 | name: "conv4_1"
297 | type: "Convolution"
298 | bottom: "pool3_stage1"
299 | top: "conv4_1"
300 | param {
301 | lr_mult: 1.0
302 | decay_mult: 1
303 | }
304 | param {
305 | lr_mult: 2.0
306 | decay_mult: 0
307 | }
308 | convolution_param {
309 | num_output: 512
310 | pad: 1
311 | kernel_size: 3
312 | weight_filler {
313 | type: "gaussian"
314 | std: 0.01
315 | }
316 | bias_filler {
317 | type: "constant"
318 | }
319 | }
320 | }
321 | layer {
322 | name: "relu4_1"
323 | type: "ReLU"
324 | bottom: "conv4_1"
325 | top: "conv4_1"
326 | }
327 | layer {
328 | name: "conv4_2"
329 | type: "Convolution"
330 | bottom: "conv4_1"
331 | top: "conv4_2"
332 | param {
333 | lr_mult: 1.0
334 | decay_mult: 1
335 | }
336 | param {
337 | lr_mult: 2.0
338 | decay_mult: 0
339 | }
340 | convolution_param {
341 | num_output: 512
342 | pad: 1
343 | kernel_size: 3
344 | weight_filler {
345 | type: "gaussian"
346 | std: 0.01
347 | }
348 | bias_filler {
349 | type: "constant"
350 | }
351 | }
352 | }
353 | layer {
354 | name: "relu4_2"
355 | type: "ReLU"
356 | bottom: "conv4_2"
357 | top: "conv4_2"
358 | }
359 | layer {
360 | name: "conv4_3_CPM"
361 | type: "Convolution"
362 | bottom: "conv4_2"
363 | top: "conv4_3_CPM"
364 | param {
365 | lr_mult: 1.0
366 | decay_mult: 1
367 | }
368 | param {
369 | lr_mult: 2.0
370 | decay_mult: 0
371 | }
372 | convolution_param {
373 | num_output: 256
374 | pad: 1
375 | kernel_size: 3
376 | weight_filler {
377 | type: "gaussian"
378 | std: 0.01
379 | }
380 | bias_filler {
381 | type: "constant"
382 | }
383 | }
384 | }
385 | layer {
386 | name: "relu4_3_CPM"
387 | type: "ReLU"
388 | bottom: "conv4_3_CPM"
389 | top: "conv4_3_CPM"
390 | }
391 | layer {
392 | name: "conv4_4_CPM"
393 | type: "Convolution"
394 | bottom: "conv4_3_CPM"
395 | top: "conv4_4_CPM"
396 | param {
397 | lr_mult: 1.0
398 | decay_mult: 1
399 | }
400 | param {
401 | lr_mult: 2.0
402 | decay_mult: 0
403 | }
404 | convolution_param {
405 | num_output: 128
406 | pad: 1
407 | kernel_size: 3
408 | weight_filler {
409 | type: "gaussian"
410 | std: 0.01
411 | }
412 | bias_filler {
413 | type: "constant"
414 | }
415 | }
416 | }
417 | layer {
418 | name: "relu4_4_CPM"
419 | type: "ReLU"
420 | bottom: "conv4_4_CPM"
421 | top: "conv4_4_CPM"
422 | }
423 | layer {
424 | name: "conv5_1_CPM_L1"
425 | type: "Convolution"
426 | bottom: "conv4_4_CPM"
427 | top: "conv5_1_CPM_L1"
428 | param {
429 | lr_mult: 1.0
430 | decay_mult: 1
431 | }
432 | param {
433 | lr_mult: 2.0
434 | decay_mult: 0
435 | }
436 | convolution_param {
437 | num_output: 128
438 | pad: 1
439 | kernel_size: 3
440 | weight_filler {
441 | type: "gaussian"
442 | std: 0.01
443 | }
444 | bias_filler {
445 | type: "constant"
446 | }
447 | }
448 | }
449 | layer {
450 | name: "relu5_1_CPM_L1"
451 | type: "ReLU"
452 | bottom: "conv5_1_CPM_L1"
453 | top: "conv5_1_CPM_L1"
454 | }
455 | layer {
456 | name: "conv5_1_CPM_L2"
457 | type: "Convolution"
458 | bottom: "conv4_4_CPM"
459 | top: "conv5_1_CPM_L2"
460 | param {
461 | lr_mult: 1.0
462 | decay_mult: 1
463 | }
464 | param {
465 | lr_mult: 2.0
466 | decay_mult: 0
467 | }
468 | convolution_param {
469 | num_output: 128
470 | pad: 1
471 | kernel_size: 3
472 | weight_filler {
473 | type: "gaussian"
474 | std: 0.01
475 | }
476 | bias_filler {
477 | type: "constant"
478 | }
479 | }
480 | }
481 | layer {
482 | name: "relu5_1_CPM_L2"
483 | type: "ReLU"
484 | bottom: "conv5_1_CPM_L2"
485 | top: "conv5_1_CPM_L2"
486 | }
487 | layer {
488 | name: "conv5_2_CPM_L1"
489 | type: "Convolution"
490 | bottom: "conv5_1_CPM_L1"
491 | top: "conv5_2_CPM_L1"
492 | param {
493 | lr_mult: 1.0
494 | decay_mult: 1
495 | }
496 | param {
497 | lr_mult: 2.0
498 | decay_mult: 0
499 | }
500 | convolution_param {
501 | num_output: 128
502 | pad: 1
503 | kernel_size: 3
504 | weight_filler {
505 | type: "gaussian"
506 | std: 0.01
507 | }
508 | bias_filler {
509 | type: "constant"
510 | }
511 | }
512 | }
513 | layer {
514 | name: "relu5_2_CPM_L1"
515 | type: "ReLU"
516 | bottom: "conv5_2_CPM_L1"
517 | top: "conv5_2_CPM_L1"
518 | }
519 | layer {
520 | name: "conv5_2_CPM_L2"
521 | type: "Convolution"
522 | bottom: "conv5_1_CPM_L2"
523 | top: "conv5_2_CPM_L2"
524 | param {
525 | lr_mult: 1.0
526 | decay_mult: 1
527 | }
528 | param {
529 | lr_mult: 2.0
530 | decay_mult: 0
531 | }
532 | convolution_param {
533 | num_output: 128
534 | pad: 1
535 | kernel_size: 3
536 | weight_filler {
537 | type: "gaussian"
538 | std: 0.01
539 | }
540 | bias_filler {
541 | type: "constant"
542 | }
543 | }
544 | }
545 | layer {
546 | name: "relu5_2_CPM_L2"
547 | type: "ReLU"
548 | bottom: "conv5_2_CPM_L2"
549 | top: "conv5_2_CPM_L2"
550 | }
551 | layer {
552 | name: "conv5_3_CPM_L1"
553 | type: "Convolution"
554 | bottom: "conv5_2_CPM_L1"
555 | top: "conv5_3_CPM_L1"
556 | param {
557 | lr_mult: 1.0
558 | decay_mult: 1
559 | }
560 | param {
561 | lr_mult: 2.0
562 | decay_mult: 0
563 | }
564 | convolution_param {
565 | num_output: 128
566 | pad: 1
567 | kernel_size: 3
568 | weight_filler {
569 | type: "gaussian"
570 | std: 0.01
571 | }
572 | bias_filler {
573 | type: "constant"
574 | }
575 | }
576 | }
577 | layer {
578 | name: "relu5_3_CPM_L1"
579 | type: "ReLU"
580 | bottom: "conv5_3_CPM_L1"
581 | top: "conv5_3_CPM_L1"
582 | }
583 | layer {
584 | name: "conv5_3_CPM_L2"
585 | type: "Convolution"
586 | bottom: "conv5_2_CPM_L2"
587 | top: "conv5_3_CPM_L2"
588 | param {
589 | lr_mult: 1.0
590 | decay_mult: 1
591 | }
592 | param {
593 | lr_mult: 2.0
594 | decay_mult: 0
595 | }
596 | convolution_param {
597 | num_output: 128
598 | pad: 1
599 | kernel_size: 3
600 | weight_filler {
601 | type: "gaussian"
602 | std: 0.01
603 | }
604 | bias_filler {
605 | type: "constant"
606 | }
607 | }
608 | }
609 | layer {
610 | name: "relu5_3_CPM_L2"
611 | type: "ReLU"
612 | bottom: "conv5_3_CPM_L2"
613 | top: "conv5_3_CPM_L2"
614 | }
615 | layer {
616 | name: "conv5_4_CPM_L1"
617 | type: "Convolution"
618 | bottom: "conv5_3_CPM_L1"
619 | top: "conv5_4_CPM_L1"
620 | param {
621 | lr_mult: 1.0
622 | decay_mult: 1
623 | }
624 | param {
625 | lr_mult: 2.0
626 | decay_mult: 0
627 | }
628 | convolution_param {
629 | num_output: 512
630 | pad: 0
631 | kernel_size: 1
632 | weight_filler {
633 | type: "gaussian"
634 | std: 0.01
635 | }
636 | bias_filler {
637 | type: "constant"
638 | }
639 | }
640 | }
641 | layer {
642 | name: "relu5_4_CPM_L1"
643 | type: "ReLU"
644 | bottom: "conv5_4_CPM_L1"
645 | top: "conv5_4_CPM_L1"
646 | }
647 | layer {
648 | name: "conv5_4_CPM_L2"
649 | type: "Convolution"
650 | bottom: "conv5_3_CPM_L2"
651 | top: "conv5_4_CPM_L2"
652 | param {
653 | lr_mult: 1.0
654 | decay_mult: 1
655 | }
656 | param {
657 | lr_mult: 2.0
658 | decay_mult: 0
659 | }
660 | convolution_param {
661 | num_output: 512
662 | pad: 0
663 | kernel_size: 1
664 | weight_filler {
665 | type: "gaussian"
666 | std: 0.01
667 | }
668 | bias_filler {
669 | type: "constant"
670 | }
671 | }
672 | }
673 | layer {
674 | name: "relu5_4_CPM_L2"
675 | type: "ReLU"
676 | bottom: "conv5_4_CPM_L2"
677 | top: "conv5_4_CPM_L2"
678 | }
679 | layer {
680 | name: "conv5_5_CPM_L1"
681 | type: "Convolution"
682 | bottom: "conv5_4_CPM_L1"
683 | top: "conv5_5_CPM_L1"
684 | param {
685 | lr_mult: 1.0
686 | decay_mult: 1
687 | }
688 | param {
689 | lr_mult: 2.0
690 | decay_mult: 0
691 | }
692 | convolution_param {
693 | num_output: 28
694 | pad: 0
695 | kernel_size: 1
696 | weight_filler {
697 | type: "gaussian"
698 | std: 0.01
699 | }
700 | bias_filler {
701 | type: "constant"
702 | }
703 | }
704 | }
705 | layer {
706 | name: "conv5_5_CPM_L2"
707 | type: "Convolution"
708 | bottom: "conv5_4_CPM_L2"
709 | top: "conv5_5_CPM_L2"
710 | param {
711 | lr_mult: 1.0
712 | decay_mult: 1
713 | }
714 | param {
715 | lr_mult: 2.0
716 | decay_mult: 0
717 | }
718 | convolution_param {
719 | num_output: 16
720 | pad: 0
721 | kernel_size: 1
722 | weight_filler {
723 | type: "gaussian"
724 | std: 0.01
725 | }
726 | bias_filler {
727 | type: "constant"
728 | }
729 | }
730 | }
731 | layer {
732 | name: "concat_stage2"
733 | type: "Concat"
734 | bottom: "conv5_5_CPM_L1"
735 | bottom: "conv5_5_CPM_L2"
736 | bottom: "conv4_4_CPM"
737 | top: "concat_stage2"
738 | concat_param {
739 | axis: 1
740 | }
741 | }
742 | layer {
743 | name: "Mconv1_stage2_L1"
744 | type: "Convolution"
745 | bottom: "concat_stage2"
746 | top: "Mconv1_stage2_L1"
747 | param {
748 | lr_mult: 4.0
749 | decay_mult: 1
750 | }
751 | param {
752 | lr_mult: 8.0
753 | decay_mult: 0
754 | }
755 | convolution_param {
756 | num_output: 128
757 | pad: 3
758 | kernel_size: 7
759 | weight_filler {
760 | type: "gaussian"
761 | std: 0.01
762 | }
763 | bias_filler {
764 | type: "constant"
765 | }
766 | }
767 | }
768 | layer {
769 | name: "Mrelu1_stage2_L1"
770 | type: "ReLU"
771 | bottom: "Mconv1_stage2_L1"
772 | top: "Mconv1_stage2_L1"
773 | }
774 | layer {
775 | name: "Mconv1_stage2_L2"
776 | type: "Convolution"
777 | bottom: "concat_stage2"
778 | top: "Mconv1_stage2_L2"
779 | param {
780 | lr_mult: 4.0
781 | decay_mult: 1
782 | }
783 | param {
784 | lr_mult: 8.0
785 | decay_mult: 0
786 | }
787 | convolution_param {
788 | num_output: 128
789 | pad: 3
790 | kernel_size: 7
791 | weight_filler {
792 | type: "gaussian"
793 | std: 0.01
794 | }
795 | bias_filler {
796 | type: "constant"
797 | }
798 | }
799 | }
800 | layer {
801 | name: "Mrelu1_stage2_L2"
802 | type: "ReLU"
803 | bottom: "Mconv1_stage2_L2"
804 | top: "Mconv1_stage2_L2"
805 | }
806 | layer {
807 | name: "Mconv2_stage2_L1"
808 | type: "Convolution"
809 | bottom: "Mconv1_stage2_L1"
810 | top: "Mconv2_stage2_L1"
811 | param {
812 | lr_mult: 4.0
813 | decay_mult: 1
814 | }
815 | param {
816 | lr_mult: 8.0
817 | decay_mult: 0
818 | }
819 | convolution_param {
820 | num_output: 128
821 | pad: 3
822 | kernel_size: 7
823 | weight_filler {
824 | type: "gaussian"
825 | std: 0.01
826 | }
827 | bias_filler {
828 | type: "constant"
829 | }
830 | }
831 | }
832 | layer {
833 | name: "Mrelu2_stage2_L1"
834 | type: "ReLU"
835 | bottom: "Mconv2_stage2_L1"
836 | top: "Mconv2_stage2_L1"
837 | }
838 | layer {
839 | name: "Mconv2_stage2_L2"
840 | type: "Convolution"
841 | bottom: "Mconv1_stage2_L2"
842 | top: "Mconv2_stage2_L2"
843 | param {
844 | lr_mult: 4.0
845 | decay_mult: 1
846 | }
847 | param {
848 | lr_mult: 8.0
849 | decay_mult: 0
850 | }
851 | convolution_param {
852 | num_output: 128
853 | pad: 3
854 | kernel_size: 7
855 | weight_filler {
856 | type: "gaussian"
857 | std: 0.01
858 | }
859 | bias_filler {
860 | type: "constant"
861 | }
862 | }
863 | }
864 | layer {
865 | name: "Mrelu2_stage2_L2"
866 | type: "ReLU"
867 | bottom: "Mconv2_stage2_L2"
868 | top: "Mconv2_stage2_L2"
869 | }
870 | layer {
871 | name: "Mconv3_stage2_L1"
872 | type: "Convolution"
873 | bottom: "Mconv2_stage2_L1"
874 | top: "Mconv3_stage2_L1"
875 | param {
876 | lr_mult: 4.0
877 | decay_mult: 1
878 | }
879 | param {
880 | lr_mult: 8.0
881 | decay_mult: 0
882 | }
883 | convolution_param {
884 | num_output: 128
885 | pad: 3
886 | kernel_size: 7
887 | weight_filler {
888 | type: "gaussian"
889 | std: 0.01
890 | }
891 | bias_filler {
892 | type: "constant"
893 | }
894 | }
895 | }
896 | layer {
897 | name: "Mrelu3_stage2_L1"
898 | type: "ReLU"
899 | bottom: "Mconv3_stage2_L1"
900 | top: "Mconv3_stage2_L1"
901 | }
902 | layer {
903 | name: "Mconv3_stage2_L2"
904 | type: "Convolution"
905 | bottom: "Mconv2_stage2_L2"
906 | top: "Mconv3_stage2_L2"
907 | param {
908 | lr_mult: 4.0
909 | decay_mult: 1
910 | }
911 | param {
912 | lr_mult: 8.0
913 | decay_mult: 0
914 | }
915 | convolution_param {
916 | num_output: 128
917 | pad: 3
918 | kernel_size: 7
919 | weight_filler {
920 | type: "gaussian"
921 | std: 0.01
922 | }
923 | bias_filler {
924 | type: "constant"
925 | }
926 | }
927 | }
928 | layer {
929 | name: "Mrelu3_stage2_L2"
930 | type: "ReLU"
931 | bottom: "Mconv3_stage2_L2"
932 | top: "Mconv3_stage2_L2"
933 | }
934 | layer {
935 | name: "Mconv4_stage2_L1"
936 | type: "Convolution"
937 | bottom: "Mconv3_stage2_L1"
938 | top: "Mconv4_stage2_L1"
939 | param {
940 | lr_mult: 4.0
941 | decay_mult: 1
942 | }
943 | param {
944 | lr_mult: 8.0
945 | decay_mult: 0
946 | }
947 | convolution_param {
948 | num_output: 128
949 | pad: 3
950 | kernel_size: 7
951 | weight_filler {
952 | type: "gaussian"
953 | std: 0.01
954 | }
955 | bias_filler {
956 | type: "constant"
957 | }
958 | }
959 | }
960 | layer {
961 | name: "Mrelu4_stage2_L1"
962 | type: "ReLU"
963 | bottom: "Mconv4_stage2_L1"
964 | top: "Mconv4_stage2_L1"
965 | }
966 | layer {
967 | name: "Mconv4_stage2_L2"
968 | type: "Convolution"
969 | bottom: "Mconv3_stage2_L2"
970 | top: "Mconv4_stage2_L2"
971 | param {
972 | lr_mult: 4.0
973 | decay_mult: 1
974 | }
975 | param {
976 | lr_mult: 8.0
977 | decay_mult: 0
978 | }
979 | convolution_param {
980 | num_output: 128
981 | pad: 3
982 | kernel_size: 7
983 | weight_filler {
984 | type: "gaussian"
985 | std: 0.01
986 | }
987 | bias_filler {
988 | type: "constant"
989 | }
990 | }
991 | }
992 | layer {
993 | name: "Mrelu4_stage2_L2"
994 | type: "ReLU"
995 | bottom: "Mconv4_stage2_L2"
996 | top: "Mconv4_stage2_L2"
997 | }
998 | layer {
999 | name: "Mconv5_stage2_L1"
1000 | type: "Convolution"
1001 | bottom: "Mconv4_stage2_L1"
1002 | top: "Mconv5_stage2_L1"
1003 | param {
1004 | lr_mult: 4.0
1005 | decay_mult: 1
1006 | }
1007 | param {
1008 | lr_mult: 8.0
1009 | decay_mult: 0
1010 | }
1011 | convolution_param {
1012 | num_output: 128
1013 | pad: 3
1014 | kernel_size: 7
1015 | weight_filler {
1016 | type: "gaussian"
1017 | std: 0.01
1018 | }
1019 | bias_filler {
1020 | type: "constant"
1021 | }
1022 | }
1023 | }
1024 | layer {
1025 | name: "Mrelu5_stage2_L1"
1026 | type: "ReLU"
1027 | bottom: "Mconv5_stage2_L1"
1028 | top: "Mconv5_stage2_L1"
1029 | }
1030 | layer {
1031 | name: "Mconv5_stage2_L2"
1032 | type: "Convolution"
1033 | bottom: "Mconv4_stage2_L2"
1034 | top: "Mconv5_stage2_L2"
1035 | param {
1036 | lr_mult: 4.0
1037 | decay_mult: 1
1038 | }
1039 | param {
1040 | lr_mult: 8.0
1041 | decay_mult: 0
1042 | }
1043 | convolution_param {
1044 | num_output: 128
1045 | pad: 3
1046 | kernel_size: 7
1047 | weight_filler {
1048 | type: "gaussian"
1049 | std: 0.01
1050 | }
1051 | bias_filler {
1052 | type: "constant"
1053 | }
1054 | }
1055 | }
1056 | layer {
1057 | name: "Mrelu5_stage2_L2"
1058 | type: "ReLU"
1059 | bottom: "Mconv5_stage2_L2"
1060 | top: "Mconv5_stage2_L2"
1061 | }
1062 | layer {
1063 | name: "Mconv6_stage2_L1"
1064 | type: "Convolution"
1065 | bottom: "Mconv5_stage2_L1"
1066 | top: "Mconv6_stage2_L1"
1067 | param {
1068 | lr_mult: 4.0
1069 | decay_mult: 1
1070 | }
1071 | param {
1072 | lr_mult: 8.0
1073 | decay_mult: 0
1074 | }
1075 | convolution_param {
1076 | num_output: 128
1077 | pad: 0
1078 | kernel_size: 1
1079 | weight_filler {
1080 | type: "gaussian"
1081 | std: 0.01
1082 | }
1083 | bias_filler {
1084 | type: "constant"
1085 | }
1086 | }
1087 | }
1088 | layer {
1089 | name: "Mrelu6_stage2_L1"
1090 | type: "ReLU"
1091 | bottom: "Mconv6_stage2_L1"
1092 | top: "Mconv6_stage2_L1"
1093 | }
1094 | layer {
1095 | name: "Mconv6_stage2_L2"
1096 | type: "Convolution"
1097 | bottom: "Mconv5_stage2_L2"
1098 | top: "Mconv6_stage2_L2"
1099 | param {
1100 | lr_mult: 4.0
1101 | decay_mult: 1
1102 | }
1103 | param {
1104 | lr_mult: 8.0
1105 | decay_mult: 0
1106 | }
1107 | convolution_param {
1108 | num_output: 128
1109 | pad: 0
1110 | kernel_size: 1
1111 | weight_filler {
1112 | type: "gaussian"
1113 | std: 0.01
1114 | }
1115 | bias_filler {
1116 | type: "constant"
1117 | }
1118 | }
1119 | }
1120 | layer {
1121 | name: "Mrelu6_stage2_L2"
1122 | type: "ReLU"
1123 | bottom: "Mconv6_stage2_L2"
1124 | top: "Mconv6_stage2_L2"
1125 | }
1126 | layer {
1127 | name: "Mconv7_stage2_L1"
1128 | type: "Convolution"
1129 | bottom: "Mconv6_stage2_L1"
1130 | top: "Mconv7_stage2_L1"
1131 | param {
1132 | lr_mult: 4.0
1133 | decay_mult: 1
1134 | }
1135 | param {
1136 | lr_mult: 8.0
1137 | decay_mult: 0
1138 | }
1139 | convolution_param {
1140 | num_output: 28
1141 | pad: 0
1142 | kernel_size: 1
1143 | weight_filler {
1144 | type: "gaussian"
1145 | std: 0.01
1146 | }
1147 | bias_filler {
1148 | type: "constant"
1149 | }
1150 | }
1151 | }
1152 | layer {
1153 | name: "Mconv7_stage2_L2"
1154 | type: "Convolution"
1155 | bottom: "Mconv6_stage2_L2"
1156 | top: "Mconv7_stage2_L2"
1157 | param {
1158 | lr_mult: 4.0
1159 | decay_mult: 1
1160 | }
1161 | param {
1162 | lr_mult: 8.0
1163 | decay_mult: 0
1164 | }
1165 | convolution_param {
1166 | num_output: 16
1167 | pad: 0
1168 | kernel_size: 1
1169 | weight_filler {
1170 | type: "gaussian"
1171 | std: 0.01
1172 | }
1173 | bias_filler {
1174 | type: "constant"
1175 | }
1176 | }
1177 | }
1178 | layer {
1179 | name: "concat_stage3"
1180 | type: "Concat"
1181 | bottom: "Mconv7_stage2_L1"
1182 | bottom: "Mconv7_stage2_L2"
1183 | bottom: "conv4_4_CPM"
1184 | top: "concat_stage3"
1185 | concat_param {
1186 | axis: 1
1187 | }
1188 | }
1189 | layer {
1190 | name: "Mconv1_stage3_L1"
1191 | type: "Convolution"
1192 | bottom: "concat_stage3"
1193 | top: "Mconv1_stage3_L1"
1194 | param {
1195 | lr_mult: 4.0
1196 | decay_mult: 1
1197 | }
1198 | param {
1199 | lr_mult: 8.0
1200 | decay_mult: 0
1201 | }
1202 | convolution_param {
1203 | num_output: 128
1204 | pad: 3
1205 | kernel_size: 7
1206 | weight_filler {
1207 | type: "gaussian"
1208 | std: 0.01
1209 | }
1210 | bias_filler {
1211 | type: "constant"
1212 | }
1213 | }
1214 | }
1215 | layer {
1216 | name: "Mrelu1_stage3_L1"
1217 | type: "ReLU"
1218 | bottom: "Mconv1_stage3_L1"
1219 | top: "Mconv1_stage3_L1"
1220 | }
1221 | layer {
1222 | name: "Mconv1_stage3_L2"
1223 | type: "Convolution"
1224 | bottom: "concat_stage3"
1225 | top: "Mconv1_stage3_L2"
1226 | param {
1227 | lr_mult: 4.0
1228 | decay_mult: 1
1229 | }
1230 | param {
1231 | lr_mult: 8.0
1232 | decay_mult: 0
1233 | }
1234 | convolution_param {
1235 | num_output: 128
1236 | pad: 3
1237 | kernel_size: 7
1238 | weight_filler {
1239 | type: "gaussian"
1240 | std: 0.01
1241 | }
1242 | bias_filler {
1243 | type: "constant"
1244 | }
1245 | }
1246 | }
1247 | layer {
1248 | name: "Mrelu1_stage3_L2"
1249 | type: "ReLU"
1250 | bottom: "Mconv1_stage3_L2"
1251 | top: "Mconv1_stage3_L2"
1252 | }
1253 | layer {
1254 | name: "Mconv2_stage3_L1"
1255 | type: "Convolution"
1256 | bottom: "Mconv1_stage3_L1"
1257 | top: "Mconv2_stage3_L1"
1258 | param {
1259 | lr_mult: 4.0
1260 | decay_mult: 1
1261 | }
1262 | param {
1263 | lr_mult: 8.0
1264 | decay_mult: 0
1265 | }
1266 | convolution_param {
1267 | num_output: 128
1268 | pad: 3
1269 | kernel_size: 7
1270 | weight_filler {
1271 | type: "gaussian"
1272 | std: 0.01
1273 | }
1274 | bias_filler {
1275 | type: "constant"
1276 | }
1277 | }
1278 | }
1279 | layer {
1280 | name: "Mrelu2_stage3_L1"
1281 | type: "ReLU"
1282 | bottom: "Mconv2_stage3_L1"
1283 | top: "Mconv2_stage3_L1"
1284 | }
1285 | layer {
1286 | name: "Mconv2_stage3_L2"
1287 | type: "Convolution"
1288 | bottom: "Mconv1_stage3_L2"
1289 | top: "Mconv2_stage3_L2"
1290 | param {
1291 | lr_mult: 4.0
1292 | decay_mult: 1
1293 | }
1294 | param {
1295 | lr_mult: 8.0
1296 | decay_mult: 0
1297 | }
1298 | convolution_param {
1299 | num_output: 128
1300 | pad: 3
1301 | kernel_size: 7
1302 | weight_filler {
1303 | type: "gaussian"
1304 | std: 0.01
1305 | }
1306 | bias_filler {
1307 | type: "constant"
1308 | }
1309 | }
1310 | }
1311 | layer {
1312 | name: "Mrelu2_stage3_L2"
1313 | type: "ReLU"
1314 | bottom: "Mconv2_stage3_L2"
1315 | top: "Mconv2_stage3_L2"
1316 | }
1317 | layer {
1318 | name: "Mconv3_stage3_L1"
1319 | type: "Convolution"
1320 | bottom: "Mconv2_stage3_L1"
1321 | top: "Mconv3_stage3_L1"
1322 | param {
1323 | lr_mult: 4.0
1324 | decay_mult: 1
1325 | }
1326 | param {
1327 | lr_mult: 8.0
1328 | decay_mult: 0
1329 | }
1330 | convolution_param {
1331 | num_output: 128
1332 | pad: 3
1333 | kernel_size: 7
1334 | weight_filler {
1335 | type: "gaussian"
1336 | std: 0.01
1337 | }
1338 | bias_filler {
1339 | type: "constant"
1340 | }
1341 | }
1342 | }
1343 | layer {
1344 | name: "Mrelu3_stage3_L1"
1345 | type: "ReLU"
1346 | bottom: "Mconv3_stage3_L1"
1347 | top: "Mconv3_stage3_L1"
1348 | }
1349 | layer {
1350 | name: "Mconv3_stage3_L2"
1351 | type: "Convolution"
1352 | bottom: "Mconv2_stage3_L2"
1353 | top: "Mconv3_stage3_L2"
1354 | param {
1355 | lr_mult: 4.0
1356 | decay_mult: 1
1357 | }
1358 | param {
1359 | lr_mult: 8.0
1360 | decay_mult: 0
1361 | }
1362 | convolution_param {
1363 | num_output: 128
1364 | pad: 3
1365 | kernel_size: 7
1366 | weight_filler {
1367 | type: "gaussian"
1368 | std: 0.01
1369 | }
1370 | bias_filler {
1371 | type: "constant"
1372 | }
1373 | }
1374 | }
1375 | layer {
1376 | name: "Mrelu3_stage3_L2"
1377 | type: "ReLU"
1378 | bottom: "Mconv3_stage3_L2"
1379 | top: "Mconv3_stage3_L2"
1380 | }
1381 | layer {
1382 | name: "Mconv4_stage3_L1"
1383 | type: "Convolution"
1384 | bottom: "Mconv3_stage3_L1"
1385 | top: "Mconv4_stage3_L1"
1386 | param {
1387 | lr_mult: 4.0
1388 | decay_mult: 1
1389 | }
1390 | param {
1391 | lr_mult: 8.0
1392 | decay_mult: 0
1393 | }
1394 | convolution_param {
1395 | num_output: 128
1396 | pad: 3
1397 | kernel_size: 7
1398 | weight_filler {
1399 | type: "gaussian"
1400 | std: 0.01
1401 | }
1402 | bias_filler {
1403 | type: "constant"
1404 | }
1405 | }
1406 | }
1407 | layer {
1408 | name: "Mrelu4_stage3_L1"
1409 | type: "ReLU"
1410 | bottom: "Mconv4_stage3_L1"
1411 | top: "Mconv4_stage3_L1"
1412 | }
1413 | layer {
1414 | name: "Mconv4_stage3_L2"
1415 | type: "Convolution"
1416 | bottom: "Mconv3_stage3_L2"
1417 | top: "Mconv4_stage3_L2"
1418 | param {
1419 | lr_mult: 4.0
1420 | decay_mult: 1
1421 | }
1422 | param {
1423 | lr_mult: 8.0
1424 | decay_mult: 0
1425 | }
1426 | convolution_param {
1427 | num_output: 128
1428 | pad: 3
1429 | kernel_size: 7
1430 | weight_filler {
1431 | type: "gaussian"
1432 | std: 0.01
1433 | }
1434 | bias_filler {
1435 | type: "constant"
1436 | }
1437 | }
1438 | }
1439 | layer {
1440 | name: "Mrelu4_stage3_L2"
1441 | type: "ReLU"
1442 | bottom: "Mconv4_stage3_L2"
1443 | top: "Mconv4_stage3_L2"
1444 | }
1445 | layer {
1446 | name: "Mconv5_stage3_L1"
1447 | type: "Convolution"
1448 | bottom: "Mconv4_stage3_L1"
1449 | top: "Mconv5_stage3_L1"
1450 | param {
1451 | lr_mult: 4.0
1452 | decay_mult: 1
1453 | }
1454 | param {
1455 | lr_mult: 8.0
1456 | decay_mult: 0
1457 | }
1458 | convolution_param {
1459 | num_output: 128
1460 | pad: 3
1461 | kernel_size: 7
1462 | weight_filler {
1463 | type: "gaussian"
1464 | std: 0.01
1465 | }
1466 | bias_filler {
1467 | type: "constant"
1468 | }
1469 | }
1470 | }
1471 | layer {
1472 | name: "Mrelu5_stage3_L1"
1473 | type: "ReLU"
1474 | bottom: "Mconv5_stage3_L1"
1475 | top: "Mconv5_stage3_L1"
1476 | }
1477 | layer {
1478 | name: "Mconv5_stage3_L2"
1479 | type: "Convolution"
1480 | bottom: "Mconv4_stage3_L2"
1481 | top: "Mconv5_stage3_L2"
1482 | param {
1483 | lr_mult: 4.0
1484 | decay_mult: 1
1485 | }
1486 | param {
1487 | lr_mult: 8.0
1488 | decay_mult: 0
1489 | }
1490 | convolution_param {
1491 | num_output: 128
1492 | pad: 3
1493 | kernel_size: 7
1494 | weight_filler {
1495 | type: "gaussian"
1496 | std: 0.01
1497 | }
1498 | bias_filler {
1499 | type: "constant"
1500 | }
1501 | }
1502 | }
1503 | layer {
1504 | name: "Mrelu5_stage3_L2"
1505 | type: "ReLU"
1506 | bottom: "Mconv5_stage3_L2"
1507 | top: "Mconv5_stage3_L2"
1508 | }
1509 | layer {
1510 | name: "Mconv6_stage3_L1"
1511 | type: "Convolution"
1512 | bottom: "Mconv5_stage3_L1"
1513 | top: "Mconv6_stage3_L1"
1514 | param {
1515 | lr_mult: 4.0
1516 | decay_mult: 1
1517 | }
1518 | param {
1519 | lr_mult: 8.0
1520 | decay_mult: 0
1521 | }
1522 | convolution_param {
1523 | num_output: 128
1524 | pad: 0
1525 | kernel_size: 1
1526 | weight_filler {
1527 | type: "gaussian"
1528 | std: 0.01
1529 | }
1530 | bias_filler {
1531 | type: "constant"
1532 | }
1533 | }
1534 | }
1535 | layer {
1536 | name: "Mrelu6_stage3_L1"
1537 | type: "ReLU"
1538 | bottom: "Mconv6_stage3_L1"
1539 | top: "Mconv6_stage3_L1"
1540 | }
1541 | layer {
1542 | name: "Mconv6_stage3_L2"
1543 | type: "Convolution"
1544 | bottom: "Mconv5_stage3_L2"
1545 | top: "Mconv6_stage3_L2"
1546 | param {
1547 | lr_mult: 4.0
1548 | decay_mult: 1
1549 | }
1550 | param {
1551 | lr_mult: 8.0
1552 | decay_mult: 0
1553 | }
1554 | convolution_param {
1555 | num_output: 128
1556 | pad: 0
1557 | kernel_size: 1
1558 | weight_filler {
1559 | type: "gaussian"
1560 | std: 0.01
1561 | }
1562 | bias_filler {
1563 | type: "constant"
1564 | }
1565 | }
1566 | }
1567 | layer {
1568 | name: "Mrelu6_stage3_L2"
1569 | type: "ReLU"
1570 | bottom: "Mconv6_stage3_L2"
1571 | top: "Mconv6_stage3_L2"
1572 | }
1573 | layer {
1574 | name: "Mconv7_stage3_L1"
1575 | type: "Convolution"
1576 | bottom: "Mconv6_stage3_L1"
1577 | top: "Mconv7_stage3_L1"
1578 | param {
1579 | lr_mult: 4.0
1580 | decay_mult: 1
1581 | }
1582 | param {
1583 | lr_mult: 8.0
1584 | decay_mult: 0
1585 | }
1586 | convolution_param {
1587 | num_output: 28
1588 | pad: 0
1589 | kernel_size: 1
1590 | weight_filler {
1591 | type: "gaussian"
1592 | std: 0.01
1593 | }
1594 | bias_filler {
1595 | type: "constant"
1596 | }
1597 | }
1598 | }
1599 | layer {
1600 | name: "Mconv7_stage3_L2"
1601 | type: "Convolution"
1602 | bottom: "Mconv6_stage3_L2"
1603 | top: "Mconv7_stage3_L2"
1604 | param {
1605 | lr_mult: 4.0
1606 | decay_mult: 1
1607 | }
1608 | param {
1609 | lr_mult: 8.0
1610 | decay_mult: 0
1611 | }
1612 | convolution_param {
1613 | num_output: 16
1614 | pad: 0
1615 | kernel_size: 1
1616 | weight_filler {
1617 | type: "gaussian"
1618 | std: 0.01
1619 | }
1620 | bias_filler {
1621 | type: "constant"
1622 | }
1623 | }
1624 | }
1625 | layer {
1626 | name: "concat_stage4"
1627 | type: "Concat"
1628 | bottom: "Mconv7_stage3_L1"
1629 | bottom: "Mconv7_stage3_L2"
1630 | bottom: "conv4_4_CPM"
1631 | top: "concat_stage4"
1632 | concat_param {
1633 | axis: 1
1634 | }
1635 | }
1636 | layer {
1637 | name: "Mconv1_stage4_L1"
1638 | type: "Convolution"
1639 | bottom: "concat_stage4"
1640 | top: "Mconv1_stage4_L1"
1641 | param {
1642 | lr_mult: 4.0
1643 | decay_mult: 1
1644 | }
1645 | param {
1646 | lr_mult: 8.0
1647 | decay_mult: 0
1648 | }
1649 | convolution_param {
1650 | num_output: 128
1651 | pad: 3
1652 | kernel_size: 7
1653 | weight_filler {
1654 | type: "gaussian"
1655 | std: 0.01
1656 | }
1657 | bias_filler {
1658 | type: "constant"
1659 | }
1660 | }
1661 | }
1662 | layer {
1663 | name: "Mrelu1_stage4_L1"
1664 | type: "ReLU"
1665 | bottom: "Mconv1_stage4_L1"
1666 | top: "Mconv1_stage4_L1"
1667 | }
1668 | layer {
1669 | name: "Mconv1_stage4_L2"
1670 | type: "Convolution"
1671 | bottom: "concat_stage4"
1672 | top: "Mconv1_stage4_L2"
1673 | param {
1674 | lr_mult: 4.0
1675 | decay_mult: 1
1676 | }
1677 | param {
1678 | lr_mult: 8.0
1679 | decay_mult: 0
1680 | }
1681 | convolution_param {
1682 | num_output: 128
1683 | pad: 3
1684 | kernel_size: 7
1685 | weight_filler {
1686 | type: "gaussian"
1687 | std: 0.01
1688 | }
1689 | bias_filler {
1690 | type: "constant"
1691 | }
1692 | }
1693 | }
1694 | layer {
1695 | name: "Mrelu1_stage4_L2"
1696 | type: "ReLU"
1697 | bottom: "Mconv1_stage4_L2"
1698 | top: "Mconv1_stage4_L2"
1699 | }
1700 | layer {
1701 | name: "Mconv2_stage4_L1"
1702 | type: "Convolution"
1703 | bottom: "Mconv1_stage4_L1"
1704 | top: "Mconv2_stage4_L1"
1705 | param {
1706 | lr_mult: 4.0
1707 | decay_mult: 1
1708 | }
1709 | param {
1710 | lr_mult: 8.0
1711 | decay_mult: 0
1712 | }
1713 | convolution_param {
1714 | num_output: 128
1715 | pad: 3
1716 | kernel_size: 7
1717 | weight_filler {
1718 | type: "gaussian"
1719 | std: 0.01
1720 | }
1721 | bias_filler {
1722 | type: "constant"
1723 | }
1724 | }
1725 | }
1726 | layer {
1727 | name: "Mrelu2_stage4_L1"
1728 | type: "ReLU"
1729 | bottom: "Mconv2_stage4_L1"
1730 | top: "Mconv2_stage4_L1"
1731 | }
1732 | layer {
1733 | name: "Mconv2_stage4_L2"
1734 | type: "Convolution"
1735 | bottom: "Mconv1_stage4_L2"
1736 | top: "Mconv2_stage4_L2"
1737 | param {
1738 | lr_mult: 4.0
1739 | decay_mult: 1
1740 | }
1741 | param {
1742 | lr_mult: 8.0
1743 | decay_mult: 0
1744 | }
1745 | convolution_param {
1746 | num_output: 128
1747 | pad: 3
1748 | kernel_size: 7
1749 | weight_filler {
1750 | type: "gaussian"
1751 | std: 0.01
1752 | }
1753 | bias_filler {
1754 | type: "constant"
1755 | }
1756 | }
1757 | }
1758 | layer {
1759 | name: "Mrelu2_stage4_L2"
1760 | type: "ReLU"
1761 | bottom: "Mconv2_stage4_L2"
1762 | top: "Mconv2_stage4_L2"
1763 | }
1764 | layer {
1765 | name: "Mconv3_stage4_L1"
1766 | type: "Convolution"
1767 | bottom: "Mconv2_stage4_L1"
1768 | top: "Mconv3_stage4_L1"
1769 | param {
1770 | lr_mult: 4.0
1771 | decay_mult: 1
1772 | }
1773 | param {
1774 | lr_mult: 8.0
1775 | decay_mult: 0
1776 | }
1777 | convolution_param {
1778 | num_output: 128
1779 | pad: 3
1780 | kernel_size: 7
1781 | weight_filler {
1782 | type: "gaussian"
1783 | std: 0.01
1784 | }
1785 | bias_filler {
1786 | type: "constant"
1787 | }
1788 | }
1789 | }
1790 | layer {
1791 | name: "Mrelu3_stage4_L1"
1792 | type: "ReLU"
1793 | bottom: "Mconv3_stage4_L1"
1794 | top: "Mconv3_stage4_L1"
1795 | }
1796 | layer {
1797 | name: "Mconv3_stage4_L2"
1798 | type: "Convolution"
1799 | bottom: "Mconv2_stage4_L2"
1800 | top: "Mconv3_stage4_L2"
1801 | param {
1802 | lr_mult: 4.0
1803 | decay_mult: 1
1804 | }
1805 | param {
1806 | lr_mult: 8.0
1807 | decay_mult: 0
1808 | }
1809 | convolution_param {
1810 | num_output: 128
1811 | pad: 3
1812 | kernel_size: 7
1813 | weight_filler {
1814 | type: "gaussian"
1815 | std: 0.01
1816 | }
1817 | bias_filler {
1818 | type: "constant"
1819 | }
1820 | }
1821 | }
1822 | layer {
1823 | name: "Mrelu3_stage4_L2"
1824 | type: "ReLU"
1825 | bottom: "Mconv3_stage4_L2"
1826 | top: "Mconv3_stage4_L2"
1827 | }
1828 | layer {
1829 | name: "Mconv4_stage4_L1"
1830 | type: "Convolution"
1831 | bottom: "Mconv3_stage4_L1"
1832 | top: "Mconv4_stage4_L1"
1833 | param {
1834 | lr_mult: 4.0
1835 | decay_mult: 1
1836 | }
1837 | param {
1838 | lr_mult: 8.0
1839 | decay_mult: 0
1840 | }
1841 | convolution_param {
1842 | num_output: 128
1843 | pad: 3
1844 | kernel_size: 7
1845 | weight_filler {
1846 | type: "gaussian"
1847 | std: 0.01
1848 | }
1849 | bias_filler {
1850 | type: "constant"
1851 | }
1852 | }
1853 | }
1854 | layer {
1855 | name: "Mrelu4_stage4_L1"
1856 | type: "ReLU"
1857 | bottom: "Mconv4_stage4_L1"
1858 | top: "Mconv4_stage4_L1"
1859 | }
1860 | layer {
1861 | name: "Mconv4_stage4_L2"
1862 | type: "Convolution"
1863 | bottom: "Mconv3_stage4_L2"
1864 | top: "Mconv4_stage4_L2"
1865 | param {
1866 | lr_mult: 4.0
1867 | decay_mult: 1
1868 | }
1869 | param {
1870 | lr_mult: 8.0
1871 | decay_mult: 0
1872 | }
1873 | convolution_param {
1874 | num_output: 128
1875 | pad: 3
1876 | kernel_size: 7
1877 | weight_filler {
1878 | type: "gaussian"
1879 | std: 0.01
1880 | }
1881 | bias_filler {
1882 | type: "constant"
1883 | }
1884 | }
1885 | }
1886 | layer {
1887 | name: "Mrelu4_stage4_L2"
1888 | type: "ReLU"
1889 | bottom: "Mconv4_stage4_L2"
1890 | top: "Mconv4_stage4_L2"
1891 | }
1892 | layer {
1893 | name: "Mconv5_stage4_L1"
1894 | type: "Convolution"
1895 | bottom: "Mconv4_stage4_L1"
1896 | top: "Mconv5_stage4_L1"
1897 | param {
1898 | lr_mult: 4.0
1899 | decay_mult: 1
1900 | }
1901 | param {
1902 | lr_mult: 8.0
1903 | decay_mult: 0
1904 | }
1905 | convolution_param {
1906 | num_output: 128
1907 | pad: 3
1908 | kernel_size: 7
1909 | weight_filler {
1910 | type: "gaussian"
1911 | std: 0.01
1912 | }
1913 | bias_filler {
1914 | type: "constant"
1915 | }
1916 | }
1917 | }
1918 | layer {
1919 | name: "Mrelu5_stage4_L1"
1920 | type: "ReLU"
1921 | bottom: "Mconv5_stage4_L1"
1922 | top: "Mconv5_stage4_L1"
1923 | }
1924 | layer {
1925 | name: "Mconv5_stage4_L2"
1926 | type: "Convolution"
1927 | bottom: "Mconv4_stage4_L2"
1928 | top: "Mconv5_stage4_L2"
1929 | param {
1930 | lr_mult: 4.0
1931 | decay_mult: 1
1932 | }
1933 | param {
1934 | lr_mult: 8.0
1935 | decay_mult: 0
1936 | }
1937 | convolution_param {
1938 | num_output: 128
1939 | pad: 3
1940 | kernel_size: 7
1941 | weight_filler {
1942 | type: "gaussian"
1943 | std: 0.01
1944 | }
1945 | bias_filler {
1946 | type: "constant"
1947 | }
1948 | }
1949 | }
1950 | layer {
1951 | name: "Mrelu5_stage4_L2"
1952 | type: "ReLU"
1953 | bottom: "Mconv5_stage4_L2"
1954 | top: "Mconv5_stage4_L2"
1955 | }
1956 | layer {
1957 | name: "Mconv6_stage4_L1"
1958 | type: "Convolution"
1959 | bottom: "Mconv5_stage4_L1"
1960 | top: "Mconv6_stage4_L1"
1961 | param {
1962 | lr_mult: 4.0
1963 | decay_mult: 1
1964 | }
1965 | param {
1966 | lr_mult: 8.0
1967 | decay_mult: 0
1968 | }
1969 | convolution_param {
1970 | num_output: 128
1971 | pad: 0
1972 | kernel_size: 1
1973 | weight_filler {
1974 | type: "gaussian"
1975 | std: 0.01
1976 | }
1977 | bias_filler {
1978 | type: "constant"
1979 | }
1980 | }
1981 | }
1982 | layer {
1983 | name: "Mrelu6_stage4_L1"
1984 | type: "ReLU"
1985 | bottom: "Mconv6_stage4_L1"
1986 | top: "Mconv6_stage4_L1"
1987 | }
1988 | layer {
1989 | name: "Mconv6_stage4_L2"
1990 | type: "Convolution"
1991 | bottom: "Mconv5_stage4_L2"
1992 | top: "Mconv6_stage4_L2"
1993 | param {
1994 | lr_mult: 4.0
1995 | decay_mult: 1
1996 | }
1997 | param {
1998 | lr_mult: 8.0
1999 | decay_mult: 0
2000 | }
2001 | convolution_param {
2002 | num_output: 128
2003 | pad: 0
2004 | kernel_size: 1
2005 | weight_filler {
2006 | type: "gaussian"
2007 | std: 0.01
2008 | }
2009 | bias_filler {
2010 | type: "constant"
2011 | }
2012 | }
2013 | }
2014 | layer {
2015 | name: "Mrelu6_stage4_L2"
2016 | type: "ReLU"
2017 | bottom: "Mconv6_stage4_L2"
2018 | top: "Mconv6_stage4_L2"
2019 | }
2020 | layer {
2021 | name: "Mconv7_stage4_L1"
2022 | type: "Convolution"
2023 | bottom: "Mconv6_stage4_L1"
2024 | top: "Mconv7_stage4_L1"
2025 | param {
2026 | lr_mult: 4.0
2027 | decay_mult: 1
2028 | }
2029 | param {
2030 | lr_mult: 8.0
2031 | decay_mult: 0
2032 | }
2033 | convolution_param {
2034 | num_output: 28
2035 | pad: 0
2036 | kernel_size: 1
2037 | weight_filler {
2038 | type: "gaussian"
2039 | std: 0.01
2040 | }
2041 | bias_filler {
2042 | type: "constant"
2043 | }
2044 | }
2045 | }
2046 | layer {
2047 | name: "Mconv7_stage4_L2"
2048 | type: "Convolution"
2049 | bottom: "Mconv6_stage4_L2"
2050 | top: "Mconv7_stage4_L2"
2051 | param {
2052 | lr_mult: 4.0
2053 | decay_mult: 1
2054 | }
2055 | param {
2056 | lr_mult: 8.0
2057 | decay_mult: 0
2058 | }
2059 | convolution_param {
2060 | num_output: 16
2061 | pad: 0
2062 | kernel_size: 1
2063 | weight_filler {
2064 | type: "gaussian"
2065 | std: 0.01
2066 | }
2067 | bias_filler {
2068 | type: "constant"
2069 | }
2070 | }
2071 | }
2072 | layer {
2073 | name: "concat_stage7"
2074 | type: "Concat"
2075 | bottom: "Mconv7_stage4_L2"
2076 | bottom: "Mconv7_stage4_L1"
2077 | top: "net_output"
2078 | concat_param {
2079 | axis: 1
2080 | }
2081 | }
2082 |
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/model/mpi/pose_deploy_linevec_faster_4_stages.prototxt:
--------------------------------------------------------------------------------
1 | input: "image"
2 | input_dim: 1
3 | input_dim: 3
4 | input_dim: 1 # This value will be defined at runtime
5 | input_dim: 1 # This value will be defined at runtime
6 | layer {
7 | name: "conv1_1"
8 | type: "Convolution"
9 | bottom: "image"
10 | top: "conv1_1"
11 | param {
12 | lr_mult: 1.0
13 | decay_mult: 1
14 | }
15 | param {
16 | lr_mult: 2.0
17 | decay_mult: 0
18 | }
19 | convolution_param {
20 | num_output: 64
21 | pad: 1
22 | kernel_size: 3
23 | weight_filler {
24 | type: "gaussian"
25 | std: 0.01
26 | }
27 | bias_filler {
28 | type: "constant"
29 | }
30 | }
31 | }
32 | layer {
33 | name: "relu1_1"
34 | type: "ReLU"
35 | bottom: "conv1_1"
36 | top: "conv1_1"
37 | }
38 | layer {
39 | name: "conv1_2"
40 | type: "Convolution"
41 | bottom: "conv1_1"
42 | top: "conv1_2"
43 | param {
44 | lr_mult: 1.0
45 | decay_mult: 1
46 | }
47 | param {
48 | lr_mult: 2.0
49 | decay_mult: 0
50 | }
51 | convolution_param {
52 | num_output: 64
53 | pad: 1
54 | kernel_size: 3
55 | weight_filler {
56 | type: "gaussian"
57 | std: 0.01
58 | }
59 | bias_filler {
60 | type: "constant"
61 | }
62 | }
63 | }
64 | layer {
65 | name: "relu1_2"
66 | type: "ReLU"
67 | bottom: "conv1_2"
68 | top: "conv1_2"
69 | }
70 | layer {
71 | name: "pool1_stage1"
72 | type: "Pooling"
73 | bottom: "conv1_2"
74 | top: "pool1_stage1"
75 | pooling_param {
76 | pool: MAX
77 | kernel_size: 2
78 | stride: 2
79 | }
80 | }
81 | layer {
82 | name: "conv2_1"
83 | type: "Convolution"
84 | bottom: "pool1_stage1"
85 | top: "conv2_1"
86 | param {
87 | lr_mult: 1.0
88 | decay_mult: 1
89 | }
90 | param {
91 | lr_mult: 2.0
92 | decay_mult: 0
93 | }
94 | convolution_param {
95 | num_output: 128
96 | pad: 1
97 | kernel_size: 3
98 | weight_filler {
99 | type: "gaussian"
100 | std: 0.01
101 | }
102 | bias_filler {
103 | type: "constant"
104 | }
105 | }
106 | }
107 | layer {
108 | name: "relu2_1"
109 | type: "ReLU"
110 | bottom: "conv2_1"
111 | top: "conv2_1"
112 | }
113 | layer {
114 | name: "conv2_2"
115 | type: "Convolution"
116 | bottom: "conv2_1"
117 | top: "conv2_2"
118 | param {
119 | lr_mult: 1.0
120 | decay_mult: 1
121 | }
122 | param {
123 | lr_mult: 2.0
124 | decay_mult: 0
125 | }
126 | convolution_param {
127 | num_output: 128
128 | pad: 1
129 | kernel_size: 3
130 | weight_filler {
131 | type: "gaussian"
132 | std: 0.01
133 | }
134 | bias_filler {
135 | type: "constant"
136 | }
137 | }
138 | }
139 | layer {
140 | name: "relu2_2"
141 | type: "ReLU"
142 | bottom: "conv2_2"
143 | top: "conv2_2"
144 | }
145 | layer {
146 | name: "pool2_stage1"
147 | type: "Pooling"
148 | bottom: "conv2_2"
149 | top: "pool2_stage1"
150 | pooling_param {
151 | pool: MAX
152 | kernel_size: 2
153 | stride: 2
154 | }
155 | }
156 | layer {
157 | name: "conv3_1"
158 | type: "Convolution"
159 | bottom: "pool2_stage1"
160 | top: "conv3_1"
161 | param {
162 | lr_mult: 1.0
163 | decay_mult: 1
164 | }
165 | param {
166 | lr_mult: 2.0
167 | decay_mult: 0
168 | }
169 | convolution_param {
170 | num_output: 256
171 | pad: 1
172 | kernel_size: 3
173 | weight_filler {
174 | type: "gaussian"
175 | std: 0.01
176 | }
177 | bias_filler {
178 | type: "constant"
179 | }
180 | }
181 | }
182 | layer {
183 | name: "relu3_1"
184 | type: "ReLU"
185 | bottom: "conv3_1"
186 | top: "conv3_1"
187 | }
188 | layer {
189 | name: "conv3_2"
190 | type: "Convolution"
191 | bottom: "conv3_1"
192 | top: "conv3_2"
193 | param {
194 | lr_mult: 1.0
195 | decay_mult: 1
196 | }
197 | param {
198 | lr_mult: 2.0
199 | decay_mult: 0
200 | }
201 | convolution_param {
202 | num_output: 256
203 | pad: 1
204 | kernel_size: 3
205 | weight_filler {
206 | type: "gaussian"
207 | std: 0.01
208 | }
209 | bias_filler {
210 | type: "constant"
211 | }
212 | }
213 | }
214 | layer {
215 | name: "relu3_2"
216 | type: "ReLU"
217 | bottom: "conv3_2"
218 | top: "conv3_2"
219 | }
220 | layer {
221 | name: "conv3_3"
222 | type: "Convolution"
223 | bottom: "conv3_2"
224 | top: "conv3_3"
225 | param {
226 | lr_mult: 1.0
227 | decay_mult: 1
228 | }
229 | param {
230 | lr_mult: 2.0
231 | decay_mult: 0
232 | }
233 | convolution_param {
234 | num_output: 256
235 | pad: 1
236 | kernel_size: 3
237 | weight_filler {
238 | type: "gaussian"
239 | std: 0.01
240 | }
241 | bias_filler {
242 | type: "constant"
243 | }
244 | }
245 | }
246 | layer {
247 | name: "relu3_3"
248 | type: "ReLU"
249 | bottom: "conv3_3"
250 | top: "conv3_3"
251 | }
252 | layer {
253 | name: "conv3_4"
254 | type: "Convolution"
255 | bottom: "conv3_3"
256 | top: "conv3_4"
257 | param {
258 | lr_mult: 1.0
259 | decay_mult: 1
260 | }
261 | param {
262 | lr_mult: 2.0
263 | decay_mult: 0
264 | }
265 | convolution_param {
266 | num_output: 256
267 | pad: 1
268 | kernel_size: 3
269 | weight_filler {
270 | type: "gaussian"
271 | std: 0.01
272 | }
273 | bias_filler {
274 | type: "constant"
275 | }
276 | }
277 | }
278 | layer {
279 | name: "relu3_4"
280 | type: "ReLU"
281 | bottom: "conv3_4"
282 | top: "conv3_4"
283 | }
284 | layer {
285 | name: "pool3_stage1"
286 | type: "Pooling"
287 | bottom: "conv3_4"
288 | top: "pool3_stage1"
289 | pooling_param {
290 | pool: MAX
291 | kernel_size: 2
292 | stride: 2
293 | }
294 | }
295 | layer {
296 | name: "conv4_1"
297 | type: "Convolution"
298 | bottom: "pool3_stage1"
299 | top: "conv4_1"
300 | param {
301 | lr_mult: 1.0
302 | decay_mult: 1
303 | }
304 | param {
305 | lr_mult: 2.0
306 | decay_mult: 0
307 | }
308 | convolution_param {
309 | num_output: 512
310 | pad: 1
311 | kernel_size: 3
312 | weight_filler {
313 | type: "gaussian"
314 | std: 0.01
315 | }
316 | bias_filler {
317 | type: "constant"
318 | }
319 | }
320 | }
321 | layer {
322 | name: "relu4_1"
323 | type: "ReLU"
324 | bottom: "conv4_1"
325 | top: "conv4_1"
326 | }
327 | layer {
328 | name: "conv4_2"
329 | type: "Convolution"
330 | bottom: "conv4_1"
331 | top: "conv4_2"
332 | param {
333 | lr_mult: 1.0
334 | decay_mult: 1
335 | }
336 | param {
337 | lr_mult: 2.0
338 | decay_mult: 0
339 | }
340 | convolution_param {
341 | num_output: 512
342 | pad: 1
343 | kernel_size: 3
344 | weight_filler {
345 | type: "gaussian"
346 | std: 0.01
347 | }
348 | bias_filler {
349 | type: "constant"
350 | }
351 | }
352 | }
353 | layer {
354 | name: "relu4_2"
355 | type: "ReLU"
356 | bottom: "conv4_2"
357 | top: "conv4_2"
358 | }
359 | layer {
360 | name: "conv4_3_CPM"
361 | type: "Convolution"
362 | bottom: "conv4_2"
363 | top: "conv4_3_CPM"
364 | param {
365 | lr_mult: 1.0
366 | decay_mult: 1
367 | }
368 | param {
369 | lr_mult: 2.0
370 | decay_mult: 0
371 | }
372 | convolution_param {
373 | num_output: 256
374 | pad: 1
375 | kernel_size: 3
376 | weight_filler {
377 | type: "gaussian"
378 | std: 0.01
379 | }
380 | bias_filler {
381 | type: "constant"
382 | }
383 | }
384 | }
385 | layer {
386 | name: "relu4_3_CPM"
387 | type: "ReLU"
388 | bottom: "conv4_3_CPM"
389 | top: "conv4_3_CPM"
390 | }
391 | layer {
392 | name: "conv4_4_CPM"
393 | type: "Convolution"
394 | bottom: "conv4_3_CPM"
395 | top: "conv4_4_CPM"
396 | param {
397 | lr_mult: 1.0
398 | decay_mult: 1
399 | }
400 | param {
401 | lr_mult: 2.0
402 | decay_mult: 0
403 | }
404 | convolution_param {
405 | num_output: 128
406 | pad: 1
407 | kernel_size: 3
408 | weight_filler {
409 | type: "gaussian"
410 | std: 0.01
411 | }
412 | bias_filler {
413 | type: "constant"
414 | }
415 | }
416 | }
417 | layer {
418 | name: "relu4_4_CPM"
419 | type: "ReLU"
420 | bottom: "conv4_4_CPM"
421 | top: "conv4_4_CPM"
422 | }
423 | layer {
424 | name: "conv5_1_CPM_L1"
425 | type: "Convolution"
426 | bottom: "conv4_4_CPM"
427 | top: "conv5_1_CPM_L1"
428 | param {
429 | lr_mult: 1.0
430 | decay_mult: 1
431 | }
432 | param {
433 | lr_mult: 2.0
434 | decay_mult: 0
435 | }
436 | convolution_param {
437 | num_output: 128
438 | pad: 1
439 | kernel_size: 3
440 | weight_filler {
441 | type: "gaussian"
442 | std: 0.01
443 | }
444 | bias_filler {
445 | type: "constant"
446 | }
447 | }
448 | }
449 | layer {
450 | name: "relu5_1_CPM_L1"
451 | type: "ReLU"
452 | bottom: "conv5_1_CPM_L1"
453 | top: "conv5_1_CPM_L1"
454 | }
455 | layer {
456 | name: "conv5_1_CPM_L2"
457 | type: "Convolution"
458 | bottom: "conv4_4_CPM"
459 | top: "conv5_1_CPM_L2"
460 | param {
461 | lr_mult: 1.0
462 | decay_mult: 1
463 | }
464 | param {
465 | lr_mult: 2.0
466 | decay_mult: 0
467 | }
468 | convolution_param {
469 | num_output: 128
470 | pad: 1
471 | kernel_size: 3
472 | weight_filler {
473 | type: "gaussian"
474 | std: 0.01
475 | }
476 | bias_filler {
477 | type: "constant"
478 | }
479 | }
480 | }
481 | layer {
482 | name: "relu5_1_CPM_L2"
483 | type: "ReLU"
484 | bottom: "conv5_1_CPM_L2"
485 | top: "conv5_1_CPM_L2"
486 | }
487 | layer {
488 | name: "conv5_2_CPM_L1"
489 | type: "Convolution"
490 | bottom: "conv5_1_CPM_L1"
491 | top: "conv5_2_CPM_L1"
492 | param {
493 | lr_mult: 1.0
494 | decay_mult: 1
495 | }
496 | param {
497 | lr_mult: 2.0
498 | decay_mult: 0
499 | }
500 | convolution_param {
501 | num_output: 128
502 | pad: 1
503 | kernel_size: 3
504 | weight_filler {
505 | type: "gaussian"
506 | std: 0.01
507 | }
508 | bias_filler {
509 | type: "constant"
510 | }
511 | }
512 | }
513 | layer {
514 | name: "relu5_2_CPM_L1"
515 | type: "ReLU"
516 | bottom: "conv5_2_CPM_L1"
517 | top: "conv5_2_CPM_L1"
518 | }
519 | layer {
520 | name: "conv5_2_CPM_L2"
521 | type: "Convolution"
522 | bottom: "conv5_1_CPM_L2"
523 | top: "conv5_2_CPM_L2"
524 | param {
525 | lr_mult: 1.0
526 | decay_mult: 1
527 | }
528 | param {
529 | lr_mult: 2.0
530 | decay_mult: 0
531 | }
532 | convolution_param {
533 | num_output: 128
534 | pad: 1
535 | kernel_size: 3
536 | weight_filler {
537 | type: "gaussian"
538 | std: 0.01
539 | }
540 | bias_filler {
541 | type: "constant"
542 | }
543 | }
544 | }
545 | layer {
546 | name: "relu5_2_CPM_L2"
547 | type: "ReLU"
548 | bottom: "conv5_2_CPM_L2"
549 | top: "conv5_2_CPM_L2"
550 | }
551 | layer {
552 | name: "conv5_3_CPM_L1"
553 | type: "Convolution"
554 | bottom: "conv5_2_CPM_L1"
555 | top: "conv5_3_CPM_L1"
556 | param {
557 | lr_mult: 1.0
558 | decay_mult: 1
559 | }
560 | param {
561 | lr_mult: 2.0
562 | decay_mult: 0
563 | }
564 | convolution_param {
565 | num_output: 128
566 | pad: 1
567 | kernel_size: 3
568 | weight_filler {
569 | type: "gaussian"
570 | std: 0.01
571 | }
572 | bias_filler {
573 | type: "constant"
574 | }
575 | }
576 | }
577 | layer {
578 | name: "relu5_3_CPM_L1"
579 | type: "ReLU"
580 | bottom: "conv5_3_CPM_L1"
581 | top: "conv5_3_CPM_L1"
582 | }
583 | layer {
584 | name: "conv5_3_CPM_L2"
585 | type: "Convolution"
586 | bottom: "conv5_2_CPM_L2"
587 | top: "conv5_3_CPM_L2"
588 | param {
589 | lr_mult: 1.0
590 | decay_mult: 1
591 | }
592 | param {
593 | lr_mult: 2.0
594 | decay_mult: 0
595 | }
596 | convolution_param {
597 | num_output: 128
598 | pad: 1
599 | kernel_size: 3
600 | weight_filler {
601 | type: "gaussian"
602 | std: 0.01
603 | }
604 | bias_filler {
605 | type: "constant"
606 | }
607 | }
608 | }
609 | layer {
610 | name: "relu5_3_CPM_L2"
611 | type: "ReLU"
612 | bottom: "conv5_3_CPM_L2"
613 | top: "conv5_3_CPM_L2"
614 | }
615 | layer {
616 | name: "conv5_4_CPM_L1"
617 | type: "Convolution"
618 | bottom: "conv5_3_CPM_L1"
619 | top: "conv5_4_CPM_L1"
620 | param {
621 | lr_mult: 1.0
622 | decay_mult: 1
623 | }
624 | param {
625 | lr_mult: 2.0
626 | decay_mult: 0
627 | }
628 | convolution_param {
629 | num_output: 512
630 | pad: 0
631 | kernel_size: 1
632 | weight_filler {
633 | type: "gaussian"
634 | std: 0.01
635 | }
636 | bias_filler {
637 | type: "constant"
638 | }
639 | }
640 | }
641 | layer {
642 | name: "relu5_4_CPM_L1"
643 | type: "ReLU"
644 | bottom: "conv5_4_CPM_L1"
645 | top: "conv5_4_CPM_L1"
646 | }
647 | layer {
648 | name: "conv5_4_CPM_L2"
649 | type: "Convolution"
650 | bottom: "conv5_3_CPM_L2"
651 | top: "conv5_4_CPM_L2"
652 | param {
653 | lr_mult: 1.0
654 | decay_mult: 1
655 | }
656 | param {
657 | lr_mult: 2.0
658 | decay_mult: 0
659 | }
660 | convolution_param {
661 | num_output: 512
662 | pad: 0
663 | kernel_size: 1
664 | weight_filler {
665 | type: "gaussian"
666 | std: 0.01
667 | }
668 | bias_filler {
669 | type: "constant"
670 | }
671 | }
672 | }
673 | layer {
674 | name: "relu5_4_CPM_L2"
675 | type: "ReLU"
676 | bottom: "conv5_4_CPM_L2"
677 | top: "conv5_4_CPM_L2"
678 | }
679 | layer {
680 | name: "conv5_5_CPM_L1"
681 | type: "Convolution"
682 | bottom: "conv5_4_CPM_L1"
683 | top: "conv5_5_CPM_L1"
684 | param {
685 | lr_mult: 1.0
686 | decay_mult: 1
687 | }
688 | param {
689 | lr_mult: 2.0
690 | decay_mult: 0
691 | }
692 | convolution_param {
693 | num_output: 28
694 | pad: 0
695 | kernel_size: 1
696 | weight_filler {
697 | type: "gaussian"
698 | std: 0.01
699 | }
700 | bias_filler {
701 | type: "constant"
702 | }
703 | }
704 | }
705 | layer {
706 | name: "conv5_5_CPM_L2"
707 | type: "Convolution"
708 | bottom: "conv5_4_CPM_L2"
709 | top: "conv5_5_CPM_L2"
710 | param {
711 | lr_mult: 1.0
712 | decay_mult: 1
713 | }
714 | param {
715 | lr_mult: 2.0
716 | decay_mult: 0
717 | }
718 | convolution_param {
719 | num_output: 16
720 | pad: 0
721 | kernel_size: 1
722 | weight_filler {
723 | type: "gaussian"
724 | std: 0.01
725 | }
726 | bias_filler {
727 | type: "constant"
728 | }
729 | }
730 | }
731 | layer {
732 | name: "concat_stage2"
733 | type: "Concat"
734 | bottom: "conv5_5_CPM_L1"
735 | bottom: "conv5_5_CPM_L2"
736 | bottom: "conv4_4_CPM"
737 | top: "concat_stage2"
738 | concat_param {
739 | axis: 1
740 | }
741 | }
742 | layer {
743 | name: "Mconv1_stage2_L1"
744 | type: "Convolution"
745 | bottom: "concat_stage2"
746 | top: "Mconv1_stage2_L1"
747 | param {
748 | lr_mult: 4.0
749 | decay_mult: 1
750 | }
751 | param {
752 | lr_mult: 8.0
753 | decay_mult: 0
754 | }
755 | convolution_param {
756 | num_output: 128
757 | pad: 3
758 | kernel_size: 7
759 | weight_filler {
760 | type: "gaussian"
761 | std: 0.01
762 | }
763 | bias_filler {
764 | type: "constant"
765 | }
766 | }
767 | }
768 | layer {
769 | name: "Mrelu1_stage2_L1"
770 | type: "ReLU"
771 | bottom: "Mconv1_stage2_L1"
772 | top: "Mconv1_stage2_L1"
773 | }
774 | layer {
775 | name: "Mconv1_stage2_L2"
776 | type: "Convolution"
777 | bottom: "concat_stage2"
778 | top: "Mconv1_stage2_L2"
779 | param {
780 | lr_mult: 4.0
781 | decay_mult: 1
782 | }
783 | param {
784 | lr_mult: 8.0
785 | decay_mult: 0
786 | }
787 | convolution_param {
788 | num_output: 128
789 | pad: 3
790 | kernel_size: 7
791 | weight_filler {
792 | type: "gaussian"
793 | std: 0.01
794 | }
795 | bias_filler {
796 | type: "constant"
797 | }
798 | }
799 | }
800 | layer {
801 | name: "Mrelu1_stage2_L2"
802 | type: "ReLU"
803 | bottom: "Mconv1_stage2_L2"
804 | top: "Mconv1_stage2_L2"
805 | }
806 | layer {
807 | name: "Mconv2_stage2_L1"
808 | type: "Convolution"
809 | bottom: "Mconv1_stage2_L1"
810 | top: "Mconv2_stage2_L1"
811 | param {
812 | lr_mult: 4.0
813 | decay_mult: 1
814 | }
815 | param {
816 | lr_mult: 8.0
817 | decay_mult: 0
818 | }
819 | convolution_param {
820 | num_output: 128
821 | pad: 3
822 | kernel_size: 7
823 | weight_filler {
824 | type: "gaussian"
825 | std: 0.01
826 | }
827 | bias_filler {
828 | type: "constant"
829 | }
830 | }
831 | }
832 | layer {
833 | name: "Mrelu2_stage2_L1"
834 | type: "ReLU"
835 | bottom: "Mconv2_stage2_L1"
836 | top: "Mconv2_stage2_L1"
837 | }
838 | layer {
839 | name: "Mconv2_stage2_L2"
840 | type: "Convolution"
841 | bottom: "Mconv1_stage2_L2"
842 | top: "Mconv2_stage2_L2"
843 | param {
844 | lr_mult: 4.0
845 | decay_mult: 1
846 | }
847 | param {
848 | lr_mult: 8.0
849 | decay_mult: 0
850 | }
851 | convolution_param {
852 | num_output: 128
853 | pad: 3
854 | kernel_size: 7
855 | weight_filler {
856 | type: "gaussian"
857 | std: 0.01
858 | }
859 | bias_filler {
860 | type: "constant"
861 | }
862 | }
863 | }
864 | layer {
865 | name: "Mrelu2_stage2_L2"
866 | type: "ReLU"
867 | bottom: "Mconv2_stage2_L2"
868 | top: "Mconv2_stage2_L2"
869 | }
870 | layer {
871 | name: "Mconv3_stage2_L1"
872 | type: "Convolution"
873 | bottom: "Mconv2_stage2_L1"
874 | top: "Mconv3_stage2_L1"
875 | param {
876 | lr_mult: 4.0
877 | decay_mult: 1
878 | }
879 | param {
880 | lr_mult: 8.0
881 | decay_mult: 0
882 | }
883 | convolution_param {
884 | num_output: 128
885 | pad: 3
886 | kernel_size: 7
887 | weight_filler {
888 | type: "gaussian"
889 | std: 0.01
890 | }
891 | bias_filler {
892 | type: "constant"
893 | }
894 | }
895 | }
896 | layer {
897 | name: "Mrelu3_stage2_L1"
898 | type: "ReLU"
899 | bottom: "Mconv3_stage2_L1"
900 | top: "Mconv3_stage2_L1"
901 | }
902 | layer {
903 | name: "Mconv3_stage2_L2"
904 | type: "Convolution"
905 | bottom: "Mconv2_stage2_L2"
906 | top: "Mconv3_stage2_L2"
907 | param {
908 | lr_mult: 4.0
909 | decay_mult: 1
910 | }
911 | param {
912 | lr_mult: 8.0
913 | decay_mult: 0
914 | }
915 | convolution_param {
916 | num_output: 128
917 | pad: 3
918 | kernel_size: 7
919 | weight_filler {
920 | type: "gaussian"
921 | std: 0.01
922 | }
923 | bias_filler {
924 | type: "constant"
925 | }
926 | }
927 | }
928 | layer {
929 | name: "Mrelu3_stage2_L2"
930 | type: "ReLU"
931 | bottom: "Mconv3_stage2_L2"
932 | top: "Mconv3_stage2_L2"
933 | }
934 | layer {
935 | name: "Mconv4_stage2_L1"
936 | type: "Convolution"
937 | bottom: "Mconv3_stage2_L1"
938 | top: "Mconv4_stage2_L1"
939 | param {
940 | lr_mult: 4.0
941 | decay_mult: 1
942 | }
943 | param {
944 | lr_mult: 8.0
945 | decay_mult: 0
946 | }
947 | convolution_param {
948 | num_output: 128
949 | pad: 3
950 | kernel_size: 7
951 | weight_filler {
952 | type: "gaussian"
953 | std: 0.01
954 | }
955 | bias_filler {
956 | type: "constant"
957 | }
958 | }
959 | }
960 | layer {
961 | name: "Mrelu4_stage2_L1"
962 | type: "ReLU"
963 | bottom: "Mconv4_stage2_L1"
964 | top: "Mconv4_stage2_L1"
965 | }
966 | layer {
967 | name: "Mconv4_stage2_L2"
968 | type: "Convolution"
969 | bottom: "Mconv3_stage2_L2"
970 | top: "Mconv4_stage2_L2"
971 | param {
972 | lr_mult: 4.0
973 | decay_mult: 1
974 | }
975 | param {
976 | lr_mult: 8.0
977 | decay_mult: 0
978 | }
979 | convolution_param {
980 | num_output: 128
981 | pad: 3
982 | kernel_size: 7
983 | weight_filler {
984 | type: "gaussian"
985 | std: 0.01
986 | }
987 | bias_filler {
988 | type: "constant"
989 | }
990 | }
991 | }
992 | layer {
993 | name: "Mrelu4_stage2_L2"
994 | type: "ReLU"
995 | bottom: "Mconv4_stage2_L2"
996 | top: "Mconv4_stage2_L2"
997 | }
998 | layer {
999 | name: "Mconv5_stage2_L1"
1000 | type: "Convolution"
1001 | bottom: "Mconv4_stage2_L1"
1002 | top: "Mconv5_stage2_L1"
1003 | param {
1004 | lr_mult: 4.0
1005 | decay_mult: 1
1006 | }
1007 | param {
1008 | lr_mult: 8.0
1009 | decay_mult: 0
1010 | }
1011 | convolution_param {
1012 | num_output: 128
1013 | pad: 3
1014 | kernel_size: 7
1015 | weight_filler {
1016 | type: "gaussian"
1017 | std: 0.01
1018 | }
1019 | bias_filler {
1020 | type: "constant"
1021 | }
1022 | }
1023 | }
1024 | layer {
1025 | name: "Mrelu5_stage2_L1"
1026 | type: "ReLU"
1027 | bottom: "Mconv5_stage2_L1"
1028 | top: "Mconv5_stage2_L1"
1029 | }
1030 | layer {
1031 | name: "Mconv5_stage2_L2"
1032 | type: "Convolution"
1033 | bottom: "Mconv4_stage2_L2"
1034 | top: "Mconv5_stage2_L2"
1035 | param {
1036 | lr_mult: 4.0
1037 | decay_mult: 1
1038 | }
1039 | param {
1040 | lr_mult: 8.0
1041 | decay_mult: 0
1042 | }
1043 | convolution_param {
1044 | num_output: 128
1045 | pad: 3
1046 | kernel_size: 7
1047 | weight_filler {
1048 | type: "gaussian"
1049 | std: 0.01
1050 | }
1051 | bias_filler {
1052 | type: "constant"
1053 | }
1054 | }
1055 | }
1056 | layer {
1057 | name: "Mrelu5_stage2_L2"
1058 | type: "ReLU"
1059 | bottom: "Mconv5_stage2_L2"
1060 | top: "Mconv5_stage2_L2"
1061 | }
1062 | layer {
1063 | name: "Mconv6_stage2_L1"
1064 | type: "Convolution"
1065 | bottom: "Mconv5_stage2_L1"
1066 | top: "Mconv6_stage2_L1"
1067 | param {
1068 | lr_mult: 4.0
1069 | decay_mult: 1
1070 | }
1071 | param {
1072 | lr_mult: 8.0
1073 | decay_mult: 0
1074 | }
1075 | convolution_param {
1076 | num_output: 128
1077 | pad: 0
1078 | kernel_size: 1
1079 | weight_filler {
1080 | type: "gaussian"
1081 | std: 0.01
1082 | }
1083 | bias_filler {
1084 | type: "constant"
1085 | }
1086 | }
1087 | }
1088 | layer {
1089 | name: "Mrelu6_stage2_L1"
1090 | type: "ReLU"
1091 | bottom: "Mconv6_stage2_L1"
1092 | top: "Mconv6_stage2_L1"
1093 | }
1094 | layer {
1095 | name: "Mconv6_stage2_L2"
1096 | type: "Convolution"
1097 | bottom: "Mconv5_stage2_L2"
1098 | top: "Mconv6_stage2_L2"
1099 | param {
1100 | lr_mult: 4.0
1101 | decay_mult: 1
1102 | }
1103 | param {
1104 | lr_mult: 8.0
1105 | decay_mult: 0
1106 | }
1107 | convolution_param {
1108 | num_output: 128
1109 | pad: 0
1110 | kernel_size: 1
1111 | weight_filler {
1112 | type: "gaussian"
1113 | std: 0.01
1114 | }
1115 | bias_filler {
1116 | type: "constant"
1117 | }
1118 | }
1119 | }
1120 | layer {
1121 | name: "Mrelu6_stage2_L2"
1122 | type: "ReLU"
1123 | bottom: "Mconv6_stage2_L2"
1124 | top: "Mconv6_stage2_L2"
1125 | }
1126 | layer {
1127 | name: "Mconv7_stage2_L1"
1128 | type: "Convolution"
1129 | bottom: "Mconv6_stage2_L1"
1130 | top: "Mconv7_stage2_L1"
1131 | param {
1132 | lr_mult: 4.0
1133 | decay_mult: 1
1134 | }
1135 | param {
1136 | lr_mult: 8.0
1137 | decay_mult: 0
1138 | }
1139 | convolution_param {
1140 | num_output: 28
1141 | pad: 0
1142 | kernel_size: 1
1143 | weight_filler {
1144 | type: "gaussian"
1145 | std: 0.01
1146 | }
1147 | bias_filler {
1148 | type: "constant"
1149 | }
1150 | }
1151 | }
1152 | layer {
1153 | name: "Mconv7_stage2_L2"
1154 | type: "Convolution"
1155 | bottom: "Mconv6_stage2_L2"
1156 | top: "Mconv7_stage2_L2"
1157 | param {
1158 | lr_mult: 4.0
1159 | decay_mult: 1
1160 | }
1161 | param {
1162 | lr_mult: 8.0
1163 | decay_mult: 0
1164 | }
1165 | convolution_param {
1166 | num_output: 16
1167 | pad: 0
1168 | kernel_size: 1
1169 | weight_filler {
1170 | type: "gaussian"
1171 | std: 0.01
1172 | }
1173 | bias_filler {
1174 | type: "constant"
1175 | }
1176 | }
1177 | }
1178 | layer {
1179 | name: "concat_stage3"
1180 | type: "Concat"
1181 | bottom: "Mconv7_stage2_L1"
1182 | bottom: "Mconv7_stage2_L2"
1183 | bottom: "conv4_4_CPM"
1184 | top: "concat_stage3"
1185 | concat_param {
1186 | axis: 1
1187 | }
1188 | }
1189 | layer {
1190 | name: "Mconv1_stage3_L1"
1191 | type: "Convolution"
1192 | bottom: "concat_stage3"
1193 | top: "Mconv1_stage3_L1"
1194 | param {
1195 | lr_mult: 4.0
1196 | decay_mult: 1
1197 | }
1198 | param {
1199 | lr_mult: 8.0
1200 | decay_mult: 0
1201 | }
1202 | convolution_param {
1203 | num_output: 128
1204 | pad: 3
1205 | kernel_size: 7
1206 | weight_filler {
1207 | type: "gaussian"
1208 | std: 0.01
1209 | }
1210 | bias_filler {
1211 | type: "constant"
1212 | }
1213 | }
1214 | }
1215 | layer {
1216 | name: "Mrelu1_stage3_L1"
1217 | type: "ReLU"
1218 | bottom: "Mconv1_stage3_L1"
1219 | top: "Mconv1_stage3_L1"
1220 | }
1221 | layer {
1222 | name: "Mconv1_stage3_L2"
1223 | type: "Convolution"
1224 | bottom: "concat_stage3"
1225 | top: "Mconv1_stage3_L2"
1226 | param {
1227 | lr_mult: 4.0
1228 | decay_mult: 1
1229 | }
1230 | param {
1231 | lr_mult: 8.0
1232 | decay_mult: 0
1233 | }
1234 | convolution_param {
1235 | num_output: 128
1236 | pad: 3
1237 | kernel_size: 7
1238 | weight_filler {
1239 | type: "gaussian"
1240 | std: 0.01
1241 | }
1242 | bias_filler {
1243 | type: "constant"
1244 | }
1245 | }
1246 | }
1247 | layer {
1248 | name: "Mrelu1_stage3_L2"
1249 | type: "ReLU"
1250 | bottom: "Mconv1_stage3_L2"
1251 | top: "Mconv1_stage3_L2"
1252 | }
1253 | layer {
1254 | name: "Mconv2_stage3_L1"
1255 | type: "Convolution"
1256 | bottom: "Mconv1_stage3_L1"
1257 | top: "Mconv2_stage3_L1"
1258 | param {
1259 | lr_mult: 4.0
1260 | decay_mult: 1
1261 | }
1262 | param {
1263 | lr_mult: 8.0
1264 | decay_mult: 0
1265 | }
1266 | convolution_param {
1267 | num_output: 128
1268 | pad: 3
1269 | kernel_size: 7
1270 | weight_filler {
1271 | type: "gaussian"
1272 | std: 0.01
1273 | }
1274 | bias_filler {
1275 | type: "constant"
1276 | }
1277 | }
1278 | }
1279 | layer {
1280 | name: "Mrelu2_stage3_L1"
1281 | type: "ReLU"
1282 | bottom: "Mconv2_stage3_L1"
1283 | top: "Mconv2_stage3_L1"
1284 | }
1285 | layer {
1286 | name: "Mconv2_stage3_L2"
1287 | type: "Convolution"
1288 | bottom: "Mconv1_stage3_L2"
1289 | top: "Mconv2_stage3_L2"
1290 | param {
1291 | lr_mult: 4.0
1292 | decay_mult: 1
1293 | }
1294 | param {
1295 | lr_mult: 8.0
1296 | decay_mult: 0
1297 | }
1298 | convolution_param {
1299 | num_output: 128
1300 | pad: 3
1301 | kernel_size: 7
1302 | weight_filler {
1303 | type: "gaussian"
1304 | std: 0.01
1305 | }
1306 | bias_filler {
1307 | type: "constant"
1308 | }
1309 | }
1310 | }
1311 | layer {
1312 | name: "Mrelu2_stage3_L2"
1313 | type: "ReLU"
1314 | bottom: "Mconv2_stage3_L2"
1315 | top: "Mconv2_stage3_L2"
1316 | }
1317 | layer {
1318 | name: "Mconv3_stage3_L1"
1319 | type: "Convolution"
1320 | bottom: "Mconv2_stage3_L1"
1321 | top: "Mconv3_stage3_L1"
1322 | param {
1323 | lr_mult: 4.0
1324 | decay_mult: 1
1325 | }
1326 | param {
1327 | lr_mult: 8.0
1328 | decay_mult: 0
1329 | }
1330 | convolution_param {
1331 | num_output: 128
1332 | pad: 3
1333 | kernel_size: 7
1334 | weight_filler {
1335 | type: "gaussian"
1336 | std: 0.01
1337 | }
1338 | bias_filler {
1339 | type: "constant"
1340 | }
1341 | }
1342 | }
1343 | layer {
1344 | name: "Mrelu3_stage3_L1"
1345 | type: "ReLU"
1346 | bottom: "Mconv3_stage3_L1"
1347 | top: "Mconv3_stage3_L1"
1348 | }
1349 | layer {
1350 | name: "Mconv3_stage3_L2"
1351 | type: "Convolution"
1352 | bottom: "Mconv2_stage3_L2"
1353 | top: "Mconv3_stage3_L2"
1354 | param {
1355 | lr_mult: 4.0
1356 | decay_mult: 1
1357 | }
1358 | param {
1359 | lr_mult: 8.0
1360 | decay_mult: 0
1361 | }
1362 | convolution_param {
1363 | num_output: 128
1364 | pad: 3
1365 | kernel_size: 7
1366 | weight_filler {
1367 | type: "gaussian"
1368 | std: 0.01
1369 | }
1370 | bias_filler {
1371 | type: "constant"
1372 | }
1373 | }
1374 | }
1375 | layer {
1376 | name: "Mrelu3_stage3_L2"
1377 | type: "ReLU"
1378 | bottom: "Mconv3_stage3_L2"
1379 | top: "Mconv3_stage3_L2"
1380 | }
1381 | layer {
1382 | name: "Mconv4_stage3_L1"
1383 | type: "Convolution"
1384 | bottom: "Mconv3_stage3_L1"
1385 | top: "Mconv4_stage3_L1"
1386 | param {
1387 | lr_mult: 4.0
1388 | decay_mult: 1
1389 | }
1390 | param {
1391 | lr_mult: 8.0
1392 | decay_mult: 0
1393 | }
1394 | convolution_param {
1395 | num_output: 128
1396 | pad: 3
1397 | kernel_size: 7
1398 | weight_filler {
1399 | type: "gaussian"
1400 | std: 0.01
1401 | }
1402 | bias_filler {
1403 | type: "constant"
1404 | }
1405 | }
1406 | }
1407 | layer {
1408 | name: "Mrelu4_stage3_L1"
1409 | type: "ReLU"
1410 | bottom: "Mconv4_stage3_L1"
1411 | top: "Mconv4_stage3_L1"
1412 | }
1413 | layer {
1414 | name: "Mconv4_stage3_L2"
1415 | type: "Convolution"
1416 | bottom: "Mconv3_stage3_L2"
1417 | top: "Mconv4_stage3_L2"
1418 | param {
1419 | lr_mult: 4.0
1420 | decay_mult: 1
1421 | }
1422 | param {
1423 | lr_mult: 8.0
1424 | decay_mult: 0
1425 | }
1426 | convolution_param {
1427 | num_output: 128
1428 | pad: 3
1429 | kernel_size: 7
1430 | weight_filler {
1431 | type: "gaussian"
1432 | std: 0.01
1433 | }
1434 | bias_filler {
1435 | type: "constant"
1436 | }
1437 | }
1438 | }
1439 | layer {
1440 | name: "Mrelu4_stage3_L2"
1441 | type: "ReLU"
1442 | bottom: "Mconv4_stage3_L2"
1443 | top: "Mconv4_stage3_L2"
1444 | }
1445 | layer {
1446 | name: "Mconv5_stage3_L1"
1447 | type: "Convolution"
1448 | bottom: "Mconv4_stage3_L1"
1449 | top: "Mconv5_stage3_L1"
1450 | param {
1451 | lr_mult: 4.0
1452 | decay_mult: 1
1453 | }
1454 | param {
1455 | lr_mult: 8.0
1456 | decay_mult: 0
1457 | }
1458 | convolution_param {
1459 | num_output: 128
1460 | pad: 3
1461 | kernel_size: 7
1462 | weight_filler {
1463 | type: "gaussian"
1464 | std: 0.01
1465 | }
1466 | bias_filler {
1467 | type: "constant"
1468 | }
1469 | }
1470 | }
1471 | layer {
1472 | name: "Mrelu5_stage3_L1"
1473 | type: "ReLU"
1474 | bottom: "Mconv5_stage3_L1"
1475 | top: "Mconv5_stage3_L1"
1476 | }
1477 | layer {
1478 | name: "Mconv5_stage3_L2"
1479 | type: "Convolution"
1480 | bottom: "Mconv4_stage3_L2"
1481 | top: "Mconv5_stage3_L2"
1482 | param {
1483 | lr_mult: 4.0
1484 | decay_mult: 1
1485 | }
1486 | param {
1487 | lr_mult: 8.0
1488 | decay_mult: 0
1489 | }
1490 | convolution_param {
1491 | num_output: 128
1492 | pad: 3
1493 | kernel_size: 7
1494 | weight_filler {
1495 | type: "gaussian"
1496 | std: 0.01
1497 | }
1498 | bias_filler {
1499 | type: "constant"
1500 | }
1501 | }
1502 | }
1503 | layer {
1504 | name: "Mrelu5_stage3_L2"
1505 | type: "ReLU"
1506 | bottom: "Mconv5_stage3_L2"
1507 | top: "Mconv5_stage3_L2"
1508 | }
1509 | layer {
1510 | name: "Mconv6_stage3_L1"
1511 | type: "Convolution"
1512 | bottom: "Mconv5_stage3_L1"
1513 | top: "Mconv6_stage3_L1"
1514 | param {
1515 | lr_mult: 4.0
1516 | decay_mult: 1
1517 | }
1518 | param {
1519 | lr_mult: 8.0
1520 | decay_mult: 0
1521 | }
1522 | convolution_param {
1523 | num_output: 128
1524 | pad: 0
1525 | kernel_size: 1
1526 | weight_filler {
1527 | type: "gaussian"
1528 | std: 0.01
1529 | }
1530 | bias_filler {
1531 | type: "constant"
1532 | }
1533 | }
1534 | }
1535 | layer {
1536 | name: "Mrelu6_stage3_L1"
1537 | type: "ReLU"
1538 | bottom: "Mconv6_stage3_L1"
1539 | top: "Mconv6_stage3_L1"
1540 | }
1541 | layer {
1542 | name: "Mconv6_stage3_L2"
1543 | type: "Convolution"
1544 | bottom: "Mconv5_stage3_L2"
1545 | top: "Mconv6_stage3_L2"
1546 | param {
1547 | lr_mult: 4.0
1548 | decay_mult: 1
1549 | }
1550 | param {
1551 | lr_mult: 8.0
1552 | decay_mult: 0
1553 | }
1554 | convolution_param {
1555 | num_output: 128
1556 | pad: 0
1557 | kernel_size: 1
1558 | weight_filler {
1559 | type: "gaussian"
1560 | std: 0.01
1561 | }
1562 | bias_filler {
1563 | type: "constant"
1564 | }
1565 | }
1566 | }
1567 | layer {
1568 | name: "Mrelu6_stage3_L2"
1569 | type: "ReLU"
1570 | bottom: "Mconv6_stage3_L2"
1571 | top: "Mconv6_stage3_L2"
1572 | }
1573 | layer {
1574 | name: "Mconv7_stage3_L1"
1575 | type: "Convolution"
1576 | bottom: "Mconv6_stage3_L1"
1577 | top: "Mconv7_stage3_L1"
1578 | param {
1579 | lr_mult: 4.0
1580 | decay_mult: 1
1581 | }
1582 | param {
1583 | lr_mult: 8.0
1584 | decay_mult: 0
1585 | }
1586 | convolution_param {
1587 | num_output: 28
1588 | pad: 0
1589 | kernel_size: 1
1590 | weight_filler {
1591 | type: "gaussian"
1592 | std: 0.01
1593 | }
1594 | bias_filler {
1595 | type: "constant"
1596 | }
1597 | }
1598 | }
1599 | layer {
1600 | name: "Mconv7_stage3_L2"
1601 | type: "Convolution"
1602 | bottom: "Mconv6_stage3_L2"
1603 | top: "Mconv7_stage3_L2"
1604 | param {
1605 | lr_mult: 4.0
1606 | decay_mult: 1
1607 | }
1608 | param {
1609 | lr_mult: 8.0
1610 | decay_mult: 0
1611 | }
1612 | convolution_param {
1613 | num_output: 16
1614 | pad: 0
1615 | kernel_size: 1
1616 | weight_filler {
1617 | type: "gaussian"
1618 | std: 0.01
1619 | }
1620 | bias_filler {
1621 | type: "constant"
1622 | }
1623 | }
1624 | }
1625 | layer {
1626 | name: "concat_stage4"
1627 | type: "Concat"
1628 | bottom: "Mconv7_stage3_L1"
1629 | bottom: "Mconv7_stage3_L2"
1630 | bottom: "conv4_4_CPM"
1631 | top: "concat_stage4"
1632 | concat_param {
1633 | axis: 1
1634 | }
1635 | }
1636 | layer {
1637 | name: "Mconv1_stage4_L1"
1638 | type: "Convolution"
1639 | bottom: "concat_stage4"
1640 | top: "Mconv1_stage4_L1"
1641 | param {
1642 | lr_mult: 4.0
1643 | decay_mult: 1
1644 | }
1645 | param {
1646 | lr_mult: 8.0
1647 | decay_mult: 0
1648 | }
1649 | convolution_param {
1650 | num_output: 128
1651 | pad: 3
1652 | kernel_size: 7
1653 | weight_filler {
1654 | type: "gaussian"
1655 | std: 0.01
1656 | }
1657 | bias_filler {
1658 | type: "constant"
1659 | }
1660 | }
1661 | }
1662 | layer {
1663 | name: "Mrelu1_stage4_L1"
1664 | type: "ReLU"
1665 | bottom: "Mconv1_stage4_L1"
1666 | top: "Mconv1_stage4_L1"
1667 | }
1668 | layer {
1669 | name: "Mconv1_stage4_L2"
1670 | type: "Convolution"
1671 | bottom: "concat_stage4"
1672 | top: "Mconv1_stage4_L2"
1673 | param {
1674 | lr_mult: 4.0
1675 | decay_mult: 1
1676 | }
1677 | param {
1678 | lr_mult: 8.0
1679 | decay_mult: 0
1680 | }
1681 | convolution_param {
1682 | num_output: 128
1683 | pad: 3
1684 | kernel_size: 7
1685 | weight_filler {
1686 | type: "gaussian"
1687 | std: 0.01
1688 | }
1689 | bias_filler {
1690 | type: "constant"
1691 | }
1692 | }
1693 | }
1694 | layer {
1695 | name: "Mrelu1_stage4_L2"
1696 | type: "ReLU"
1697 | bottom: "Mconv1_stage4_L2"
1698 | top: "Mconv1_stage4_L2"
1699 | }
1700 | layer {
1701 | name: "Mconv2_stage4_L1"
1702 | type: "Convolution"
1703 | bottom: "Mconv1_stage4_L1"
1704 | top: "Mconv2_stage4_L1"
1705 | param {
1706 | lr_mult: 4.0
1707 | decay_mult: 1
1708 | }
1709 | param {
1710 | lr_mult: 8.0
1711 | decay_mult: 0
1712 | }
1713 | convolution_param {
1714 | num_output: 128
1715 | pad: 3
1716 | kernel_size: 7
1717 | weight_filler {
1718 | type: "gaussian"
1719 | std: 0.01
1720 | }
1721 | bias_filler {
1722 | type: "constant"
1723 | }
1724 | }
1725 | }
1726 | layer {
1727 | name: "Mrelu2_stage4_L1"
1728 | type: "ReLU"
1729 | bottom: "Mconv2_stage4_L1"
1730 | top: "Mconv2_stage4_L1"
1731 | }
1732 | layer {
1733 | name: "Mconv2_stage4_L2"
1734 | type: "Convolution"
1735 | bottom: "Mconv1_stage4_L2"
1736 | top: "Mconv2_stage4_L2"
1737 | param {
1738 | lr_mult: 4.0
1739 | decay_mult: 1
1740 | }
1741 | param {
1742 | lr_mult: 8.0
1743 | decay_mult: 0
1744 | }
1745 | convolution_param {
1746 | num_output: 128
1747 | pad: 3
1748 | kernel_size: 7
1749 | weight_filler {
1750 | type: "gaussian"
1751 | std: 0.01
1752 | }
1753 | bias_filler {
1754 | type: "constant"
1755 | }
1756 | }
1757 | }
1758 | layer {
1759 | name: "Mrelu2_stage4_L2"
1760 | type: "ReLU"
1761 | bottom: "Mconv2_stage4_L2"
1762 | top: "Mconv2_stage4_L2"
1763 | }
1764 | layer {
1765 | name: "Mconv3_stage4_L1"
1766 | type: "Convolution"
1767 | bottom: "Mconv2_stage4_L1"
1768 | top: "Mconv3_stage4_L1"
1769 | param {
1770 | lr_mult: 4.0
1771 | decay_mult: 1
1772 | }
1773 | param {
1774 | lr_mult: 8.0
1775 | decay_mult: 0
1776 | }
1777 | convolution_param {
1778 | num_output: 128
1779 | pad: 3
1780 | kernel_size: 7
1781 | weight_filler {
1782 | type: "gaussian"
1783 | std: 0.01
1784 | }
1785 | bias_filler {
1786 | type: "constant"
1787 | }
1788 | }
1789 | }
1790 | layer {
1791 | name: "Mrelu3_stage4_L1"
1792 | type: "ReLU"
1793 | bottom: "Mconv3_stage4_L1"
1794 | top: "Mconv3_stage4_L1"
1795 | }
1796 | layer {
1797 | name: "Mconv3_stage4_L2"
1798 | type: "Convolution"
1799 | bottom: "Mconv2_stage4_L2"
1800 | top: "Mconv3_stage4_L2"
1801 | param {
1802 | lr_mult: 4.0
1803 | decay_mult: 1
1804 | }
1805 | param {
1806 | lr_mult: 8.0
1807 | decay_mult: 0
1808 | }
1809 | convolution_param {
1810 | num_output: 128
1811 | pad: 3
1812 | kernel_size: 7
1813 | weight_filler {
1814 | type: "gaussian"
1815 | std: 0.01
1816 | }
1817 | bias_filler {
1818 | type: "constant"
1819 | }
1820 | }
1821 | }
1822 | layer {
1823 | name: "Mrelu3_stage4_L2"
1824 | type: "ReLU"
1825 | bottom: "Mconv3_stage4_L2"
1826 | top: "Mconv3_stage4_L2"
1827 | }
1828 | layer {
1829 | name: "Mconv4_stage4_L1"
1830 | type: "Convolution"
1831 | bottom: "Mconv3_stage4_L1"
1832 | top: "Mconv4_stage4_L1"
1833 | param {
1834 | lr_mult: 4.0
1835 | decay_mult: 1
1836 | }
1837 | param {
1838 | lr_mult: 8.0
1839 | decay_mult: 0
1840 | }
1841 | convolution_param {
1842 | num_output: 128
1843 | pad: 3
1844 | kernel_size: 7
1845 | weight_filler {
1846 | type: "gaussian"
1847 | std: 0.01
1848 | }
1849 | bias_filler {
1850 | type: "constant"
1851 | }
1852 | }
1853 | }
1854 | layer {
1855 | name: "Mrelu4_stage4_L1"
1856 | type: "ReLU"
1857 | bottom: "Mconv4_stage4_L1"
1858 | top: "Mconv4_stage4_L1"
1859 | }
1860 | layer {
1861 | name: "Mconv4_stage4_L2"
1862 | type: "Convolution"
1863 | bottom: "Mconv3_stage4_L2"
1864 | top: "Mconv4_stage4_L2"
1865 | param {
1866 | lr_mult: 4.0
1867 | decay_mult: 1
1868 | }
1869 | param {
1870 | lr_mult: 8.0
1871 | decay_mult: 0
1872 | }
1873 | convolution_param {
1874 | num_output: 128
1875 | pad: 3
1876 | kernel_size: 7
1877 | weight_filler {
1878 | type: "gaussian"
1879 | std: 0.01
1880 | }
1881 | bias_filler {
1882 | type: "constant"
1883 | }
1884 | }
1885 | }
1886 | layer {
1887 | name: "Mrelu4_stage4_L2"
1888 | type: "ReLU"
1889 | bottom: "Mconv4_stage4_L2"
1890 | top: "Mconv4_stage4_L2"
1891 | }
1892 | layer {
1893 | name: "Mconv5_stage4_L1"
1894 | type: "Convolution"
1895 | bottom: "Mconv4_stage4_L1"
1896 | top: "Mconv5_stage4_L1"
1897 | param {
1898 | lr_mult: 4.0
1899 | decay_mult: 1
1900 | }
1901 | param {
1902 | lr_mult: 8.0
1903 | decay_mult: 0
1904 | }
1905 | convolution_param {
1906 | num_output: 128
1907 | pad: 3
1908 | kernel_size: 7
1909 | weight_filler {
1910 | type: "gaussian"
1911 | std: 0.01
1912 | }
1913 | bias_filler {
1914 | type: "constant"
1915 | }
1916 | }
1917 | }
1918 | layer {
1919 | name: "Mrelu5_stage4_L1"
1920 | type: "ReLU"
1921 | bottom: "Mconv5_stage4_L1"
1922 | top: "Mconv5_stage4_L1"
1923 | }
1924 | layer {
1925 | name: "Mconv5_stage4_L2"
1926 | type: "Convolution"
1927 | bottom: "Mconv4_stage4_L2"
1928 | top: "Mconv5_stage4_L2"
1929 | param {
1930 | lr_mult: 4.0
1931 | decay_mult: 1
1932 | }
1933 | param {
1934 | lr_mult: 8.0
1935 | decay_mult: 0
1936 | }
1937 | convolution_param {
1938 | num_output: 128
1939 | pad: 3
1940 | kernel_size: 7
1941 | weight_filler {
1942 | type: "gaussian"
1943 | std: 0.01
1944 | }
1945 | bias_filler {
1946 | type: "constant"
1947 | }
1948 | }
1949 | }
1950 | layer {
1951 | name: "Mrelu5_stage4_L2"
1952 | type: "ReLU"
1953 | bottom: "Mconv5_stage4_L2"
1954 | top: "Mconv5_stage4_L2"
1955 | }
1956 | layer {
1957 | name: "Mconv6_stage4_L1"
1958 | type: "Convolution"
1959 | bottom: "Mconv5_stage4_L1"
1960 | top: "Mconv6_stage4_L1"
1961 | param {
1962 | lr_mult: 4.0
1963 | decay_mult: 1
1964 | }
1965 | param {
1966 | lr_mult: 8.0
1967 | decay_mult: 0
1968 | }
1969 | convolution_param {
1970 | num_output: 128
1971 | pad: 0
1972 | kernel_size: 1
1973 | weight_filler {
1974 | type: "gaussian"
1975 | std: 0.01
1976 | }
1977 | bias_filler {
1978 | type: "constant"
1979 | }
1980 | }
1981 | }
1982 | layer {
1983 | name: "Mrelu6_stage4_L1"
1984 | type: "ReLU"
1985 | bottom: "Mconv6_stage4_L1"
1986 | top: "Mconv6_stage4_L1"
1987 | }
1988 | layer {
1989 | name: "Mconv6_stage4_L2"
1990 | type: "Convolution"
1991 | bottom: "Mconv5_stage4_L2"
1992 | top: "Mconv6_stage4_L2"
1993 | param {
1994 | lr_mult: 4.0
1995 | decay_mult: 1
1996 | }
1997 | param {
1998 | lr_mult: 8.0
1999 | decay_mult: 0
2000 | }
2001 | convolution_param {
2002 | num_output: 128
2003 | pad: 0
2004 | kernel_size: 1
2005 | weight_filler {
2006 | type: "gaussian"
2007 | std: 0.01
2008 | }
2009 | bias_filler {
2010 | type: "constant"
2011 | }
2012 | }
2013 | }
2014 | layer {
2015 | name: "Mrelu6_stage4_L2"
2016 | type: "ReLU"
2017 | bottom: "Mconv6_stage4_L2"
2018 | top: "Mconv6_stage4_L2"
2019 | }
2020 | layer {
2021 | name: "Mconv7_stage4_L1"
2022 | type: "Convolution"
2023 | bottom: "Mconv6_stage4_L1"
2024 | top: "Mconv7_stage4_L1"
2025 | param {
2026 | lr_mult: 4.0
2027 | decay_mult: 1
2028 | }
2029 | param {
2030 | lr_mult: 8.0
2031 | decay_mult: 0
2032 | }
2033 | convolution_param {
2034 | num_output: 28
2035 | pad: 0
2036 | kernel_size: 1
2037 | weight_filler {
2038 | type: "gaussian"
2039 | std: 0.01
2040 | }
2041 | bias_filler {
2042 | type: "constant"
2043 | }
2044 | }
2045 | }
2046 | layer {
2047 | name: "Mconv7_stage4_L2"
2048 | type: "Convolution"
2049 | bottom: "Mconv6_stage4_L2"
2050 | top: "Mconv7_stage4_L2"
2051 | param {
2052 | lr_mult: 4.0
2053 | decay_mult: 1
2054 | }
2055 | param {
2056 | lr_mult: 8.0
2057 | decay_mult: 0
2058 | }
2059 | convolution_param {
2060 | num_output: 16
2061 | pad: 0
2062 | kernel_size: 1
2063 | weight_filler {
2064 | type: "gaussian"
2065 | std: 0.01
2066 | }
2067 | bias_filler {
2068 | type: "constant"
2069 | }
2070 | }
2071 | }
2072 | layer {
2073 | name: "concat_stage7"
2074 | type: "Concat"
2075 | bottom: "Mconv7_stage4_L2"
2076 | bottom: "Mconv7_stage4_L1"
2077 | top: "net_output"
2078 | concat_param {
2079 | axis: 1
2080 | }
2081 | }
2082 |
--------------------------------------------------------------------------------
/pose-estimator-using-caffemodel/test_out/README.md:
--------------------------------------------------------------------------------
1 | This directory is for saving the after-processing outputs.
2 |
--------------------------------------------------------------------------------
/test_out/CSBDGS_out.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/CSBDGS_out.jpg
--------------------------------------------------------------------------------
/test_out/Friends_out.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/Friends_out.jpg
--------------------------------------------------------------------------------
/test_out/Output-BODY_25.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/Output-BODY_25.jpg
--------------------------------------------------------------------------------
/test_out/Output-COCO.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/Output-COCO.jpg
--------------------------------------------------------------------------------
/test_out/Output-MPI.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/Output-MPI.jpg
--------------------------------------------------------------------------------
/test_out/TBBT_out.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/TBBT_out.jpg
--------------------------------------------------------------------------------
/test_out/WLWZ_out.jpg:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/WLWZ_out.jpg
--------------------------------------------------------------------------------
/test_out/webcam_out.gif:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/LZQthePlane/OpenPose-Rebuilt-Python/d7c927ec2d3dd724a8822bb58c3c6a936450eb14/test_out/webcam_out.gif
--------------------------------------------------------------------------------
