├── .devcontainer ├── Dockerfile ├── devcontainer.json └── environment.yml ├── .gitignore ├── LICENSE ├── README.md ├── datasets ├── __init__.py └── load_dataset.py ├── demo.ipynb ├── demo.py ├── evaluation ├── __init__.py ├── metrics.py └── show_correct_and_wrong_matches.py ├── feature_aggregation ├── __init__.py └── hdc.py ├── feature_extraction ├── __init__.py ├── feature_extractor.py ├── feature_extractor_cosplace.py ├── feature_extractor_eigenplaces.py ├── feature_extractor_holistic.py ├── feature_extractor_local.py └── feature_extractor_patchnetvlad.py ├── matching ├── __init__.py └── matching.py ├── output_images ├── examples_tp_fp.jpg ├── matchings.jpg └── pr_curve.jpg ├── requirements.txt └── setup.py /.devcontainer/Dockerfile: -------------------------------------------------------------------------------- 1 | # Use mambaforge as the base image 2 | FROM condaforge/mambaforge 3 | 4 | # Copy environment.yml (you may need to modify the path depending on where your file is located) 5 | COPY environment.yml ./ 6 | 7 | # Create the conda environment and install additional packages 8 | RUN mamba env update -n base -f environment.yml && \ 9 | mamba install autopep8 black yapf bandit flake8 mypy pycodestyle pydocstyle jupyter && \ 10 | mamba clean --all --yes 11 | -------------------------------------------------------------------------------- /.devcontainer/devcontainer.json: -------------------------------------------------------------------------------- 1 | { 2 | "name": "VPR Tutorial", 3 | "dockerFile": "Dockerfile", 4 | "extensions": [ 5 | "ms-python.python", 6 | "magicstack.MagicPython", 7 | "VisualStudioExptTeam.vscodeintellicode", 8 | "ms-toolsai.jupyter" 9 | ], 10 | "settings": { 11 | "terminal.integrated.shell.linux": "/bin/bash", 12 | "python.pythonPath": "/opt/conda/bin/python", 13 | "python.linting.enabled": true, 14 | "python.linting.pylintEnabled": true, 15 | "python.formatting.autopep8Path": "/opt/conda/bin/autopep8", 16 | "python.formatting.blackPath": "/opt/conda/bin/black", 17 | "python.formatting.yapfPath": "/opt/conda/bin/yapf", 18 | "python.linting.banditPath": "/opt/conda/bin/bandit", 19 | "python.linting.flake8Path": "/opt/conda/bin/flake8", 20 | "python.linting.mypyPath": "/opt/conda/bin/mypy", 21 | "python.linting.pycodestylePath": "/opt/conda/bin/pycodestyle", 22 | "python.linting.pydocstylePath": "/opt/conda/bin/pydocstyle", 23 | "python.linting.pylintPath": "/opt/conda/bin/pylint" 24 | }, 25 | "postCreateCommand": "echo 'conda activate base' >> ~/.bashrc", 26 | "forwardPorts": [5000, 8888] 27 | } 28 | -------------------------------------------------------------------------------- /.devcontainer/environment.yml: -------------------------------------------------------------------------------- 1 | name: vprtutorial 2 | channels: 3 | - conda-forge 4 | - tobiasrobotics 5 | dependencies: 6 | - python 7 | - numpy 8 | - pytorch 9 | - torchvision 10 | - patchnetvlad 11 | - natsort 12 | - tqdm 13 | - opencv 14 | - pillow 15 | - scikit-learn 16 | - faiss 17 | - matplotlib-base 18 | - tensorflow 19 | - tensorflow-hub 20 | - scikit-image 21 | - ipywidgets 22 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Ignore datasets 2 | images/ 3 | 4 | # Ignore Mac DS_Store files 5 | .DS_Store 6 | **/.DS_Store 7 | 8 | # Byte-compiled / optimized / DLL files 9 | __pycache__/ 10 | *.py[cod] 11 | *$py.class 12 | 13 | # C extensions 14 | *.so 15 | 16 | # Distribution / packaging 17 | .Python 18 | build/ 19 | develop-eggs/ 20 | dist/ 21 | downloads/ 22 | eggs/ 23 | .eggs/ 24 | lib/ 25 | lib64/ 26 | parts/ 27 | sdist/ 28 | var/ 29 | wheels/ 30 | share/python-wheels/ 31 | *.egg-info/ 32 | .installed.cfg 33 | *.egg 34 | MANIFEST 35 | 36 | # PyInstaller 37 | # Usually these files are written by a python script from a template 38 | # before PyInstaller builds the exe, so as to inject date/other infos 39 | into it. 40 | *.manifest 41 | *.spec 42 | 43 | # Installer logs 44 | pip-log.txt 45 | pip-delete-this-directory.txt 46 | 47 | # Unit test / coverage reports 48 | htmlcov/ 49 | .tox/ 50 | .nox/ 51 | .coverage 52 | .coverage.* 53 | .cache 54 | nosetests.xml 55 | coverage.xml 56 | *.cover 57 | *.py,cover 58 | .hypothesis/ 59 | .pytest_cache/ 60 | cover/ 61 | 62 | # Translations 63 | *.mo 64 | *.pot 65 | 66 | # Django stuff: 67 | *.log 68 | local_settings.py 69 | db.sqlite3 70 | db.sqlite3-journal 71 | 72 | # Flask stuff: 73 | instance/ 74 | .webassets-cache 75 | 76 | # Scrapy stuff: 77 | .scrapy 78 | 79 | # Sphinx documentation 80 | docs/_build/ 81 | 82 | # PyBuilder 83 | .pybuilder/ 84 | target/ 85 | 86 | # Jupyter Notebook 87 | .ipynb_checkpoints 88 | 89 | # IPython 90 | profile_default/ 91 | ipython_config.py 92 | 93 | # pyenv 94 | # For a library or package, you might want to ignore these files since 95 | the code is 96 | # intended to run in multiple environments; otherwise, check them in: 97 | # .python-version 98 | 99 | # pipenv 100 | # According to pypa/pipenv#598, it is recommended to include 101 | Pipfile.lock in version control. 102 | # However, in case of collaboration, if having platform-specific 103 | dependencies or dependencies 104 | # having no cross-platform support, pipenv may install dependencies that 105 | don't work, or not 106 | # install all needed dependencies. 107 | #Pipfile.lock 108 | 109 | # poetry 110 | # Similar to Pipfile.lock, it is generally recommended to include 111 | poetry.lock in version control. 112 | # This is especially recommended for binary packages to ensure 113 | reproducibility, and is more 114 | # commonly ignored for libraries. 115 | # 116 | https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 117 | #poetry.lock 118 | 119 | # pdm 120 | # Similar to Pipfile.lock, it is generally recommended to include 121 | pdm.lock in version control. 122 | #pdm.lock 123 | # pdm stores project-wide configurations in .pdm.toml, but it is 124 | recommended to not include it 125 | # in version control. 126 | # https://pdm.fming.dev/#use-with-ide 127 | .pdm.toml 128 | 129 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and 130 | github.com/pdm-project/pdm 131 | __pypackages__/ 132 | 133 | # Celery stuff 134 | celerybeat-schedule 135 | celerybeat.pid 136 | 137 | # SageMath parsed files 138 | *.sage.py 139 | 140 | # Environments 141 | .env 142 | .venv 143 | env/ 144 | venv/ 145 | ENV/ 146 | env.bak/ 147 | venv.bak/ 148 | 149 | # Spyder project settings 150 | .spyderproject 151 | .spyproject 152 | 153 | # Rope project settings 154 | .ropeproject 155 | 156 | # mkdocs documentation 157 | /site 158 | 159 | # mypy 160 | .mypy_cache/ 161 | .dmypy.json 162 | dmypy.json 163 | 164 | # Pyre type checker 165 | .pyre/ 166 | 167 | # pytype static type analyzer 168 | .pytype/ 169 | 170 | # Cython debug symbols 171 | cython_debug/ 172 | 173 | # PyCharm 174 | # JetBrains specific template is maintained in a separate 175 | JetBrains.gitignore that can 176 | # be found at 177 | https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 178 | # and can be added to the global gitignore or merged into this file. For 179 | a more nuclear 180 | # option (not recommended) you can uncomment the following to ignore the 181 | entire idea folder. 182 | #.idea/ 183 | 184 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Visual Place Recognition: A Tutorial 2 | Work in progress: This repository provides the example code from our paper "Visual Place Recognition: A Tutorial". 3 | The code performs VPR on the GardensPoint day_right--night_right dataset. Output is a plotted pr-curve, matching decisions, two examples for a true-positive and a false-positive matching, and the AUC performance, as shown below. 4 | 5 | If you use our work for your academic research, please refer to the following paper: 6 | ```bibtex 7 | @article{SchubertRAM2024ICRA2024, 8 | title={Visual Place Recognition: A Tutorial}, 9 | author={Schubert, Stefan and Neubert, Peer and Garg, Sourav and Milford, Michael and Fischer, Tobias}, 10 | journal={IEEE Robotics \& Automation Magazine}, 11 | year={2024}, 12 | doi={10.1109/MRA.2023.3310859} 13 | } 14 | ``` 15 | 16 | 17 | ## How to run the code 18 | ### Online (Using GitHub Codespaces) 19 | 20 | This repository is configured for use with [GitHub Codespaces](https://github.com/features/codespaces), a service that provides you with a fully-configured Visual Studio Code environment in the cloud, directly from your GitHub repository. 21 | 22 | To open this repository in a Codespace: 23 | 24 | 1. Click on the green "Code" button near the top-right corner of the repository page. 25 | 2. In the dropdown, select "Open with Codespaces", and then click on "+ New codespace". 26 | 3. Your Codespace will be created and will start automatically. This process may take a few minutes. 27 | 28 | Once your Codespace is ready, it will open in a new browser tab. This is a full-featured version of VS Code running in your browser, and it has access to all the files in your repository and all the tools installed in your Docker container. 29 | 30 | You can run commands in the terminal, edit files, debug code, commit changes, and do anything else you would normally do in VS Code. When you're done, you can close the browser tab, and your Codespace will automatically stop after a period of inactivity. 31 | 32 | ### Local 33 | ``` 34 | python3 demo.py 35 | ``` 36 | The GardensPoints Walking dataset will be downloaded automatically. You should get an output similar to this: 37 | ``` 38 | python3 demo.py 39 | ===== Load dataset 40 | ===== Load dataset GardensPoint day_right--night_right 41 | ===== Compute local DELF descriptors 42 | ===== Compute holistic HDC-DELF descriptors 43 | ===== Compute cosine similarities S 44 | ===== Match images 45 | ===== Evaluation 46 | 47 | ===== AUC (area under curve): 0.74 48 | ===== R@100P (maximum recall at 100% precision): 0.36 49 | ===== recall@K (R@K) -- R@1: 0.85 , R@5: 0.925 , R@10: 0.945 50 | ``` 51 | 52 | | Precision-recall curve | Matchings M | Examples for a true positive and a false positive | 53 | |:-------------------------:|:-------------------------:|:-------------------------:| 54 | |precision-recall curve P=f(R) | output_images/matchings.jpg | Examples for true positive (TP) and false positive (FP)| 55 | 56 | ## Requirements 57 | The code was tested with the library versions listed in [requirements.txt](./requirements.txt). Note that Tensorflow or PyTorch is only required if the corresponding image descriptor is used. If you use pip, simply: 58 | ```bash 59 | pip install -r requirements.txt 60 | ``` 61 | 62 | You can create a conda environment containing these libraries as follows (or use the provided [environment.yml](./.devcontainer/environment.yml)): 63 | ```bash 64 | mamba create -n vprtutorial python numpy pytorch torchvision natsort tqdm opencv pillow scikit-learn faiss matplotlib-base tensorflow tensorflow-hub tqdm scikit-image patchnetvlad -c conda-forge 65 | ``` 66 | 67 | 68 | ## List of existing open-source implementations for VPR (work in progress) 69 | [//]: # (use or rowspan to combine cells) 70 | 71 | 72 | ### Descriptors 73 | #### Holistic descriptors 74 | 75 | 76 | 77 | 78 | 79 | 80 | 81 | 82 | 83 | 84 | 85 | 86 | 87 | 88 | 89 | 90 | 91 | 92 | 93 | 94 | 95 | 96 | 97 | 98 | 99 | 100 | 101 | 102 | 103 | 104 | 105 | 106 | 107 | 108 | 109 | 110 | 111 | 112 | 113 | 114 | 115 | 116 | 117 | 118 | 119 | 120 | 121 | 122 |
AlexNetcode*paper
AMOSNetcodepaper
DELGcodepaper
DenseVLADcodepaper
HDC-DELFcodepaper
HybridNetcodepaper
NetVLADcodepaper
CosPlacecodepaper
EigenPlacescodepaper
123 | 124 | 125 | #### Local descriptors 126 | 127 | 128 | 129 | 130 | 131 | 132 | 133 | 134 | 135 | 136 | 137 | 138 | 139 | 140 | 141 | 142 | 143 | 144 | 145 | 146 | 147 | 148 | 149 | 150 | 151 | 152 | 153 | 154 | 155 | 156 | 157 | 158 | 159 |
D2-Netcodepaper
DELFcodepaper
LIFTcodepaper
Patch-NetVLADcodepaper
R2D2codepaper
SuperPointcodepaper
160 | 161 | #### Local descriptor aggregation 162 | 163 | 164 | 165 | 166 | 167 | 168 | 169 | 170 | 171 | 172 | 173 | 174 | 175 | 176 | 177 | 178 | 179 | 180 | 181 | 182 | 183 | 184 | 185 |
DBoW2codepaper
HDC (Hyperdimensional Computing)codepaper
iBoW (Incremental Bag-of-Words) / OBIndex2codepaper
VLAD (Vector of Locally Aggregated Descriptors)code*paper
186 | 187 | ### Sequence methods 188 | 189 | 190 | 191 | 192 | 193 | 194 | 195 | 196 | 197 | 198 | 199 | 200 | 201 | 202 | 203 | 204 | 205 | 206 | 207 | 208 | 209 | 210 | 211 | 212 | 213 | 214 | 215 | 216 | 217 | 218 | 219 | 220 | 221 | 222 | 223 | 224 | 225 | 226 | 227 | 228 | 229 | 230 | 231 | 232 | 233 | 234 | 235 | 236 |
Delta Descriptorscodepaper
MCNcodepaper
OpenSeqSLAMcode*paper
OpenSeqSLAM 2.0codepaper
OPRcodepaper
SeqConvcodepaper
SeqNetcodepaper
SeqVLADcodepaper
VPRcodepaper
237 | 238 | ### Misc 239 | 240 | 241 | 242 | 243 | 244 | 245 | 246 | 247 | 248 | 249 | 250 | 251 | 252 | 253 | 254 |
ICMcodepaperGraph optimization of the similarity matrix S
SuperGluecodepaperLocal descriptor matching
255 | 256 | *Third party code. Not provided by the authors. Code implements the author's idea or can be used to implement the authors idea. 257 | 258 | ## Soft Ground Truth for evaluation 259 | In the evaluation metrics, a soft ground truth matrix can be used to ignore images with a very small visual overlap to avoid penalization in recall and precision analysis (see Equation 6 in [our paper](https://ieeexplore.ieee.org/document/10261441)). This currently is only supported for `matching="multi"`. For use in single matching, please use a dilated hard ground truth directly. 260 | -------------------------------------------------------------------------------- /datasets/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/datasets/__init__.py -------------------------------------------------------------------------------- /datasets/load_dataset.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | import os 19 | import urllib.request 20 | import zipfile 21 | from glob import glob 22 | from PIL import Image 23 | import numpy as np 24 | from scipy.signal import convolve2d 25 | from typing import List, Tuple 26 | from abc import ABC, abstractmethod 27 | 28 | 29 | class Dataset(ABC): 30 | @abstractmethod 31 | def load(self) -> Tuple[List[np.ndarray], List[np.ndarray], np.ndarray, np.ndarray]: 32 | pass 33 | 34 | @abstractmethod 35 | def download(self, destination: str): 36 | pass 37 | 38 | 39 | class GardensPointDataset(Dataset): 40 | def __init__(self, destination: str = 'images/GardensPoint/'): 41 | self.destination = destination 42 | 43 | def load(self) -> Tuple[List[np.ndarray], List[np.ndarray], np.ndarray, np.ndarray]: 44 | print('===== Load dataset GardensPoint day_right--night_right') 45 | 46 | # download images if necessary 47 | if not os.path.exists(self.destination): 48 | self.download(self.destination) 49 | 50 | # load images 51 | fns_db = sorted(glob(self.destination + 'day_right/*.jpg')) 52 | fns_q = sorted(glob(self.destination + 'night_right/*.jpg')) 53 | 54 | imgs_db = [np.array(Image.open(fn)) for fn in fns_db] 55 | imgs_q = [np.array(Image.open(fn)) for fn in fns_q] 56 | 57 | # create ground truth 58 | GThard = np.eye(len(imgs_db)).astype('bool') 59 | GTsoft = convolve2d(GThard.astype('int'), 60 | np.ones((17, 1), 'int'), mode='same').astype('bool') 61 | 62 | return imgs_db, imgs_q, GThard, GTsoft 63 | 64 | def download(self, destination: str): 65 | print('===== GardensPoint dataset does not exist. Download to ' + destination + '...') 66 | 67 | fn = 'GardensPoint_Walking.zip' 68 | url = 'https://www.tu-chemnitz.de/etit/proaut/datasets/' + fn 69 | 70 | # create folders 71 | path = os.path.expanduser(destination) 72 | os.makedirs(path, exist_ok=True) 73 | 74 | # download 75 | urllib.request.urlretrieve(url, path + fn) 76 | 77 | # unzip 78 | with zipfile.ZipFile(path + fn, 'r') as zip_ref: 79 | zip_ref.extractall(destination) 80 | 81 | # remove zipfile 82 | os.remove(destination + fn) 83 | 84 | 85 | class StLuciaDataset(Dataset): 86 | def __init__(self, destination: str = 'images/StLucia_small/'): 87 | self.destination = destination 88 | 89 | def load(self) -> Tuple[List[np.ndarray], List[np.ndarray], np.ndarray, np.ndarray]: 90 | print('===== Load dataset StLucia 100909_0845--180809_1545 (small version)') 91 | 92 | # download images if necessary 93 | if not os.path.exists(self.destination): 94 | self.download(self.destination) 95 | 96 | # load images 97 | fns_db = sorted(glob(self.destination + '100909_0845/*.jpg')) 98 | fns_q = sorted(glob(self.destination + '180809_1545/*.jpg')) 99 | 100 | imgs_db = [np.array(Image.open(fn)) for fn in fns_db] 101 | imgs_q = [np.array(Image.open(fn)) for fn in fns_q] 102 | 103 | # create ground truth 104 | gt_data = np.load(self.destination + 'GT.npz') 105 | GThard = gt_data['GThard'].astype('bool') 106 | GTsoft = gt_data['GTsoft'].astype('bool') 107 | 108 | return imgs_db, imgs_q, GThard, GTsoft 109 | 110 | def download(self, destination: str): 111 | print('===== StLucia dataset does not exist. Download to ' + destination + '...') 112 | 113 | fn = 'StLucia_small.zip' 114 | url = 'https://www.tu-chemnitz.de/etit/proaut/datasets/' + fn 115 | 116 | # create folders 117 | path = os.path.expanduser(destination) 118 | os.makedirs(path, exist_ok=True) 119 | 120 | # download 121 | urllib.request.urlretrieve(url, path + fn) 122 | 123 | # unzip 124 | with zipfile.ZipFile(path + fn, 'r') as zip_ref: 125 | zip_ref.extractall(destination) 126 | 127 | # remove zipfile 128 | os.remove(destination + fn) 129 | 130 | 131 | class SFUDataset(Dataset): 132 | def __init__(self, destination: str = 'images/SFU/'): 133 | self.destination = destination 134 | 135 | def load(self) -> Tuple[List[np.ndarray], List[np.ndarray], np.ndarray, np.ndarray]: 136 | print('===== Load dataset SFU dry--jan') 137 | 138 | # download images if necessary 139 | if not os.path.exists(self.destination): 140 | self.download(self.destination) 141 | 142 | # load images 143 | fns_db = sorted(glob(self.destination + 'dry/*.jpg')) 144 | fns_q = sorted(glob(self.destination + 'jan/*.jpg')) 145 | 146 | imgs_db = [np.array(Image.open(fn)) for fn in fns_db] 147 | imgs_q = [np.array(Image.open(fn)) for fn in fns_q] 148 | 149 | # create ground truth 150 | gt_data = np.load(self.destination + 'GT.npz') 151 | GThard = gt_data['GThard'].astype('bool') 152 | GTsoft = gt_data['GTsoft'].astype('bool') 153 | 154 | return imgs_db, imgs_q, GThard, GTsoft 155 | 156 | def download(self, destination: str): 157 | print('===== SFU dataset does not exist. Download to ' + destination + '...') 158 | 159 | fn = 'SFU.zip' 160 | url = 'https://www.tu-chemnitz.de/etit/proaut/datasets/' + fn 161 | 162 | # create folders 163 | path = os.path.expanduser(destination) 164 | os.makedirs(path, exist_ok=True) 165 | 166 | # download 167 | urllib.request.urlretrieve(url, path + fn) 168 | 169 | # unzip 170 | with zipfile.ZipFile(path + fn, 'r') as zip_ref: 171 | zip_ref.extractall(destination) 172 | 173 | # remove zipfile 174 | os.remove(destination + fn) 175 | -------------------------------------------------------------------------------- /demo.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | 19 | import argparse 20 | import configparser 21 | import os 22 | 23 | from evaluation.metrics import createPR, recallAt100precision, recallAtK 24 | from evaluation import show_correct_and_wrong_matches 25 | from matching import matching 26 | from datasets.load_dataset import GardensPointDataset, StLuciaDataset, SFUDataset 27 | import numpy as np 28 | 29 | from matplotlib import pyplot as plt 30 | 31 | 32 | def main(): 33 | parser = argparse.ArgumentParser(description='Visual Place Recognition: A Tutorial. Code repository supplementing our paper.') 34 | parser.add_argument('--descriptor', type=str, default='HDC-DELF', choices=['HDC-DELF', 'AlexNet', 'NetVLAD', 'PatchNetVLAD', 'CosPlace', 'EigenPlaces', 'SAD'], help='Select descriptor (default: HDC-DELF)') 35 | parser.add_argument('--dataset', type=str, default='GardensPoint', choices=['GardensPoint', 'StLucia', 'SFU'], help='Select dataset (default: GardensPoint)') 36 | args = parser.parse_args() 37 | 38 | print('========== Start VPR with {} descriptor on dataset {}'.format(args.descriptor, args.dataset)) 39 | 40 | # load dataset 41 | print('===== Load dataset') 42 | if args.dataset == 'GardensPoint': 43 | dataset = GardensPointDataset() 44 | elif args.dataset == 'StLucia': 45 | dataset = StLuciaDataset() 46 | elif args.dataset == 'SFU': 47 | dataset = SFUDataset() 48 | else: 49 | raise ValueError('Unknown dataset: ' + args.dataset) 50 | 51 | imgs_db, imgs_q, GThard, GTsoft = dataset.load() 52 | 53 | if args.descriptor == 'HDC-DELF': 54 | from feature_extraction.feature_extractor_holistic import HDCDELF 55 | feature_extractor = HDCDELF() 56 | elif args.descriptor == 'AlexNet': 57 | from feature_extraction.feature_extractor_holistic import AlexNetConv3Extractor 58 | feature_extractor = AlexNetConv3Extractor() 59 | elif args.descriptor == 'SAD': 60 | from feature_extraction.feature_extractor_holistic import SAD 61 | feature_extractor = SAD() 62 | elif args.descriptor == 'NetVLAD' or args.descriptor == 'PatchNetVLAD': 63 | from feature_extraction.feature_extractor_patchnetvlad import PatchNetVLADFeatureExtractor 64 | from patchnetvlad.tools import PATCHNETVLAD_ROOT_DIR 65 | if args.descriptor == 'NetVLAD': 66 | configfile = os.path.join(PATCHNETVLAD_ROOT_DIR, 'configs/netvlad_extract.ini') 67 | else: 68 | configfile = os.path.join(PATCHNETVLAD_ROOT_DIR, 'configs/speed.ini') 69 | assert os.path.isfile(configfile) 70 | config = configparser.ConfigParser() 71 | config.read(configfile) 72 | feature_extractor = PatchNetVLADFeatureExtractor(config) 73 | elif args.descriptor == 'CosPlace': 74 | from feature_extraction.feature_extractor_cosplace import CosPlaceFeatureExtractor 75 | feature_extractor = CosPlaceFeatureExtractor() 76 | elif args.descriptor == 'EigenPlaces': 77 | from feature_extraction.feature_extractor_eigenplaces import EigenPlacesFeatureExtractor 78 | feature_extractor = EigenPlacesFeatureExtractor() 79 | else: 80 | raise ValueError('Unknown descriptor: ' + args.descriptor) 81 | 82 | if args.descriptor != 'PatchNetVLAD' and args.descriptor != 'SAD': 83 | print('===== Compute reference set descriptors') 84 | db_D_holistic = feature_extractor.compute_features(imgs_db) 85 | print('===== Compute query set descriptors') 86 | q_D_holistic = feature_extractor.compute_features(imgs_q) 87 | 88 | # normalize descriptors and compute S-matrix 89 | print('===== Compute cosine similarities S') 90 | db_D_holistic = db_D_holistic / np.linalg.norm(db_D_holistic , axis=1, keepdims=True) 91 | q_D_holistic = q_D_holistic / np.linalg.norm(q_D_holistic , axis=1, keepdims=True) 92 | S = np.matmul(db_D_holistic , q_D_holistic.transpose()) 93 | elif args.descriptor == 'SAD': 94 | print('===== Compute reference set descriptors') 95 | db_D_holistic = feature_extractor.compute_features(imgs_db) 96 | print('===== Compute query set descriptors') 97 | q_D_holistic = feature_extractor.compute_features(imgs_q) 98 | 99 | # compute similarity matrix S with sum of absolute differences (SAD) 100 | print('===== Compute similarities S from sum of absolute differences (SAD)') 101 | S = np.empty([len(imgs_db), len(imgs_q)], 'float32') 102 | for i in range(S.shape[0]): 103 | for j in range(S.shape[1]): 104 | diff = db_D_holistic[i]-q_D_holistic[j] 105 | dim = len(db_D_holistic[0]) - np.sum(np.isnan(diff)) 106 | diff[np.isnan(diff)] = 0 107 | S[i,j] = -np.sum(np.abs(diff)) / dim 108 | else: 109 | print('=== WARNING: The PatchNetVLAD code in this repository is not optimised and will be slow and memory consuming.') 110 | print('===== Compute reference set descriptors') 111 | db_D_holistic, db_D_patches = feature_extractor.compute_features(imgs_db) 112 | print('===== Compute query set descriptors') 113 | q_D_holistic, q_D_patches = feature_extractor.compute_features(imgs_q) 114 | # S_hol = np.matmul(db_D_holistic , q_D_holistic.transpose()) 115 | S = feature_extractor.local_matcher_from_numpy_single_scale(q_D_patches, db_D_patches) 116 | 117 | # show similarity matrix 118 | fig = plt.figure() 119 | plt.imshow(S) 120 | plt.axis('off') 121 | plt.title('Similarity matrix S') 122 | 123 | # matching decision making 124 | print('===== Match images') 125 | 126 | # best match per query -> Single-best-match VPR 127 | M1 = matching.best_match_per_query(S) 128 | 129 | # thresholding -> Multi-match VPR 130 | M2 = matching.thresholding(S, 'auto') 131 | TP = np.argwhere(M2 & GThard) # true positives 132 | FP = np.argwhere(M2 & ~GTsoft) # false positives 133 | 134 | # evaluation 135 | print('===== Evaluation') 136 | # show correct and wrong image matches 137 | show_correct_and_wrong_matches.show( 138 | imgs_db, imgs_q, TP, FP) # show random matches 139 | 140 | # show M's 141 | fig = plt.figure() 142 | ax1 = fig.add_subplot(121) 143 | ax1.imshow(M1) 144 | ax1.axis('off') 145 | ax1.set_title('Best match per query') 146 | ax2 = fig.add_subplot(122) 147 | ax2.imshow(M2) 148 | ax2.axis('off') 149 | ax2.set_title('Thresholding S>=thresh') 150 | 151 | # PR-curve 152 | P, R = createPR(S, GThard, GTsoft, matching='multi', n_thresh=100) 153 | plt.figure() 154 | plt.plot(R, P) 155 | plt.xlim(0, 1), plt.ylim(0, 1.01) 156 | plt.xlabel('Recall') 157 | plt.ylabel('Precision') 158 | plt.title('Result on GardensPoint day_right--night_right') 159 | plt.grid('on') 160 | plt.draw() 161 | 162 | # area under curve (AUC) 163 | AUC = np.trapz(P, R) 164 | print(f'\n===== AUC (area under curve): {AUC:.3f}') 165 | 166 | # maximum recall at 100% precision 167 | maxR = recallAt100precision(S, GThard, GTsoft, matching='multi', n_thresh=100) 168 | print(f'\n===== R@100P (maximum recall at 100% precision): {maxR:.2f}') 169 | 170 | # recall at K 171 | RatK = {} 172 | for K in [1, 5, 10]: 173 | RatK[K] = recallAtK(S, GThard, GTsoft, K=K) 174 | 175 | print(f'\n===== recall@K (R@K) -- R@1: {RatK[1]:.3f}, R@5: {RatK[5]:.3f}, R@10: {RatK[10]:.3f}') 176 | 177 | plt.show() 178 | 179 | 180 | if __name__ == "__main__": 181 | main() 182 | -------------------------------------------------------------------------------- /evaluation/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/evaluation/__init__.py -------------------------------------------------------------------------------- /evaluation/metrics.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | import numpy as np 19 | 20 | 21 | def createPR(S_in, GThard, GTsoft=None, matching='multi', n_thresh=100): 22 | """ 23 | Calculates the precision and recall at n_thresh equally spaced threshold values 24 | for a given similarity matrix S_in and ground truth matrices GThard and GTsoft for 25 | single-best-match VPR or multi-match VPR. 26 | 27 | The matrices S_in, GThard and GTsoft are two-dimensional and should all have the 28 | same shape. 29 | The matrices GThard and GTsoft should be binary matrices, where the entries are 30 | only zeros or ones. 31 | The matrix S_in should have continuous values between -Inf and Inf. Higher values 32 | indicate higher similarity. 33 | The string matching should be set to either "single" or "multi" for single-best- 34 | match VPR or multi-match VPR. 35 | The integer n_tresh controls the number of threshold values and should be >1. 36 | """ 37 | 38 | assert (S_in.shape == GThard.shape),"S_in and GThard must have the same shape" 39 | if GTsoft is not None: 40 | assert (S_in.shape == GTsoft.shape),"S_in and GTsoft must have the same shape" 41 | assert (S_in.ndim == 2),"S_in, GThard and GTsoft must be two-dimensional" 42 | assert (matching in ['single', 'multi']),"matching should contain one of the following strings: [single, multi]" 43 | assert (n_thresh > 1),"n_thresh must be >1" 44 | 45 | if GTsoft is not None and matching == 'single': 46 | raise ValueError( 47 | "GTSoft with single matching is not supported. " 48 | "Please use dilated hard ground truth directly. " 49 | "For more details, visit: https://github.com/stschubert/VPR_Tutorial" 50 | ) 51 | 52 | # ensure logical datatype in GT and GTsoft 53 | GT = GThard.astype('bool') 54 | if GTsoft is not None: 55 | GTsoft = GTsoft.astype('bool') 56 | 57 | # copy S and set elements that are only true in GTsoft to min(S) to ignore them during evaluation 58 | S = S_in.copy() 59 | if GTsoft is not None: 60 | S[GTsoft & ~GT] = S.min() 61 | 62 | # single-best-match or multi-match VPR 63 | if matching == 'single': 64 | # count the number of ground-truth positives (GTP) 65 | GTP = np.count_nonzero(GT.any(0)) 66 | 67 | # GT-values for best match per query (i.e., per column) 68 | GT = GT[np.argmax(S, axis=0), np.arange(GT.shape[1])] 69 | 70 | # similarities for best match per query (i.e., per column) 71 | S = np.max(S, axis=0) 72 | 73 | elif matching == 'multi': 74 | # count the number of ground-truth positives (GTP) 75 | GTP = np.count_nonzero(GT) # ground truth positives 76 | 77 | # init precision and recall vectors 78 | R = [0, ] 79 | P = [1, ] 80 | 81 | # select start and end treshold 82 | startV = S.max() # start-value for treshold 83 | endV = S.min() # end-value for treshold 84 | 85 | # iterate over different thresholds 86 | for i in np.linspace(startV, endV, n_thresh): 87 | B = S >= i # apply threshold 88 | 89 | TP = np.count_nonzero(GT & B) # true positives 90 | FP = np.count_nonzero((~GT) & B) # false positives 91 | 92 | P.append(TP / (TP + FP)) # precision 93 | R.append(TP / GTP) # recall 94 | 95 | return P, R 96 | 97 | 98 | def recallAt100precision(S_in, GThard, GTsoft=None, matching='multi', n_thresh=100): 99 | """ 100 | Calculates the maximum recall at 100% precision for a given similarity matrix S_in 101 | and ground truth matrices GThard and GTsoft for single-best-match VPR or multi-match 102 | VPR. 103 | 104 | The matrices S_in, GThard and GTsoft are two-dimensional and should all have the 105 | same shape. 106 | The matrices GThard and GTsoft should be binary matrices, where the entries are 107 | only zeros or ones. 108 | The matrix S_in should have continuous values between -Inf and Inf. Higher values 109 | indicate higher similarity. 110 | The string matching should be set to either "single" or "multi" for single-best- 111 | match VPR or multi-match VPR. 112 | The integer n_tresh controls the number of threshold values during the creation of 113 | the precision-recall curve and should be >1. 114 | """ 115 | 116 | assert (S_in.shape == GThard.shape),"S_in and GThard must have the same shape" 117 | if GTsoft is not None: 118 | assert (S_in.shape == GTsoft.shape),"S_in and GTsoft must have the same shape" 119 | assert (S_in.ndim == 2),"S_in, GThard and GTsoft must be two-dimensional" 120 | assert (matching in ['single', 'multi']),"matching should contain one of the following strings: [single, multi]" 121 | assert (n_thresh > 1),"n_thresh must be >1" 122 | 123 | if GTsoft is not None and matching == 'single': 124 | raise ValueError( 125 | "GTSoft with single matching is not supported. " 126 | "Please use dilated hard ground truth directly. " 127 | "For more details, visit: https://github.com/stschubert/VPR_Tutorial" 128 | ) 129 | 130 | # get precision-recall curve 131 | P, R = createPR(S_in, GThard, GTsoft, matching=matching, n_thresh=n_thresh) 132 | P = np.array(P) 133 | R = np.array(R) 134 | 135 | # recall values at 100% precision 136 | R = R[P==1] 137 | 138 | # maximum recall at 100% precision 139 | R = R.max() 140 | 141 | return R 142 | 143 | 144 | def recallAtK(S, GT, K=1): 145 | """ 146 | Calculates the recall@K for a given similarity matrix S and ground truth matrix GT. 147 | Note that this method does not support GTsoft - instead, please directly provide 148 | the dilated ground truth matrix as GT. 149 | 150 | The matrices S and GT are two-dimensional and should all have the same shape. 151 | The matrix GT should be binary, where the entries are only zeros or ones. 152 | The matrix S should have continuous values between -Inf and Inf. Higher values 153 | indicate higher similarity. 154 | The integer K>=1 defines the number of matching candidates that are selected and 155 | that must contain an actually matching image pair. 156 | """ 157 | 158 | assert (S.shape == GT.shape),"S and GT must have the same shape" 159 | assert (S.ndim == 2),"S and GT must be two-dimensional" 160 | assert (K >= 1),"K must be >=1" 161 | 162 | # ensure logical datatype in GT 163 | GT = GT.astype('bool') 164 | 165 | # discard all query images without an actually matching database image 166 | j = GT.sum(0) > 0 # columns with matches 167 | S = S[:,j] # select columns with a match 168 | GT = GT[:,j] # select columns with a match 169 | 170 | # select K highest similarities 171 | i = S.argsort(0)[-K:,:] 172 | j = np.tile(np.arange(i.shape[1]), [K, 1]) 173 | GT = GT[i, j] 174 | 175 | # recall@K 176 | RatK = np.sum(GT.sum(0) > 0) / GT.shape[1] 177 | 178 | return RatK 179 | -------------------------------------------------------------------------------- /evaluation/show_correct_and_wrong_matches.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | from matplotlib import pyplot as plt 19 | import numpy as np 20 | from skimage.transform import resize 21 | from typing import Tuple, List, Optional 22 | 23 | 24 | def add_frame(img_in: np.ndarray, color: Tuple[int, int, int]) -> np.ndarray: 25 | """ 26 | Adds a colored frame around an input image. 27 | 28 | Args: 29 | img_in (np.ndarray): A three-dimensional array representing the input image (height, width, channels). 30 | color (Tuple[int, int, int]): A tuple of three integers representing the RGB color of the frame. 31 | 32 | Returns: 33 | np.ndarray: A three-dimensional array representing the image with the added frame. 34 | """ 35 | img = img_in.copy() 36 | 37 | w = int(np.round(0.01*img.shape[1])) 38 | 39 | # pad left-right 40 | pad_lr = np.tile(np.uint8(color).reshape(1, 1, 3), (img.shape[0], w, 1)) 41 | img = np.concatenate([pad_lr, img, pad_lr], axis=1) 42 | 43 | # pad top-bottom 44 | pad_tb = np.tile(np.uint8(color).reshape(1, 1, 3), (w, img.shape[1], 1)) 45 | img = np.concatenate([pad_tb, img, pad_tb], axis=0) 46 | 47 | return img 48 | 49 | 50 | def show( 51 | db_imgs: List[np.ndarray], 52 | q_imgs: List[np.ndarray], 53 | TP: np.ndarray, 54 | FP: np.ndarray, 55 | M: Optional[np.ndarray] = None, 56 | ) -> None: 57 | """ 58 | Displays a visual comparison of true positive and false positive image pairs 59 | from a database and query set. Optionally, a similarity matrix can be included. 60 | 61 | Args: 62 | db_imgs (List[np.ndarray]): A list of 3D arrays representing the database images (height, width, channels). 63 | q_imgs (List[np.ndarray]): A list of 3D arrays representing the query images (height, width, channels). 64 | TP (np.ndarray): A two-dimensional array containing the indices of true positive pairs. 65 | FP (np.ndarray): A two-dimensional array containing the indices of false positive pairs. 66 | M (Optional[np.ndarray], optional): A two-dimensional array representing the similarity matrix. Defaults to None. 67 | 68 | Returns: 69 | None: This function displays the comparison result using matplotlib.pyplot but does not return any value. 70 | """ 71 | # true positive TP 72 | if(len(TP) == 0): 73 | print('No true positives found.') 74 | return 75 | 76 | idx_tp = np.random.permutation(len(TP))[:1] 77 | 78 | db_tp = db_imgs[int(TP[idx_tp, 0])] 79 | q_tp = q_imgs[int(TP[idx_tp, 1])] 80 | 81 | if db_tp.shape != q_tp.shape: 82 | q_tp = resize(q_tp.copy(), db_tp.shape, anti_aliasing=True) 83 | q_tp = np.uint8(q_tp*255) 84 | 85 | img = add_frame(np.concatenate([db_tp, q_tp], axis=1), [119, 172, 48]) 86 | 87 | # false positive FP 88 | try: 89 | idx_fp = np.random.permutation(len(FP))[:1] 90 | 91 | db_fp = db_imgs[int(FP[idx_fp, 0])] 92 | q_fp = q_imgs[int(FP[idx_fp, 1])] 93 | 94 | if db_fp.shape != q_fp.shape: 95 | q_fp = resize(q_fp.copy(), db_fp.shape, anti_aliasing=True) 96 | q_fp = np.uint8(q_fp*255) 97 | 98 | img_fp = add_frame(np.concatenate( 99 | [db_fp, q_fp], axis=1), [162, 20, 47]) 100 | img = np.concatenate([img, img_fp], axis=0) 101 | except: 102 | pass 103 | 104 | # concat M 105 | if M is not None: 106 | M = resize(M.copy(), (img.shape[0], img.shape[0])) 107 | M = np.uint8(M.astype('float32')*255) 108 | M = np.tile(np.expand_dims(M, -1), (1, 1, 3)) 109 | img = np.concatenate([M, img], axis=1) 110 | 111 | # show 112 | plt.figure() 113 | plt.imshow(img) 114 | plt.axis('off') 115 | plt.title('Examples for correct and wrong matches from S>=thresh') 116 | -------------------------------------------------------------------------------- /feature_aggregation/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/feature_aggregation/__init__.py -------------------------------------------------------------------------------- /feature_aggregation/hdc.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | import numpy as np 19 | from scipy.linalg import orth 20 | from typing import Union, Dict, List, Tuple 21 | 22 | 23 | class HDC(object): 24 | """ 25 | A class for implementing Hyperdimensional Computing (HDC). 26 | """ 27 | 28 | def __init__(self, Ds: List[Dict[str, np.ndarray]], nDims: int = 4096, nFeat: int = 200, nX: int = 5, nY: int = 7): 29 | """ 30 | Initializes the HDC object with the given parameters. 31 | 32 | Args: 33 | Ds (List[Dict[str, np.ndarray]]): A list of dictionaries containing 34 | descriptors. 35 | nDims (int, optional): The number of dimensions for the HDC vectors. 36 | Defaults to 4096. 37 | nFeat (int, optional): The number of features. Defaults to 200. 38 | nX (int, optional): The number of HDC vectors for the X-axis. Defaults 39 | to 5. 40 | nY (int, optional): The number of HDC vectors for the Y-axis. Defaults 41 | to 7. 42 | """ 43 | indim = Ds[0]['descriptors'].shape[1] 44 | self.nDims = nDims 45 | self.nFeat = nFeat 46 | self.nX = nX 47 | self.nY = nY 48 | self.Ds = Ds 49 | 50 | # random projection matrix for descriptors 51 | rng = np.random.default_rng(123) 52 | self.Proj = rng.standard_normal((indim, self.nDims), 'float32') 53 | self.Proj = orth(self.Proj.transpose()).transpose() 54 | 55 | # hdc vectors for poses 56 | rng = np.random.default_rng(456) 57 | self.X = (1. - 2. * (rng.random((nX, nDims), 'float32') > 0.5) 58 | ).astype('float32') 59 | self.posX = np.linspace(0., 1., nX) 60 | 61 | rng = np.random.default_rng(789) 62 | self.Y = (1. - 2. * (rng.random((nY, nDims), 'float32') > 0.5) 63 | ).astype('float32') 64 | self.posY = np.linspace(0., 1., nY) 65 | 66 | 67 | def compute_holistic(self) -> np.ndarray: 68 | """ 69 | Computes the holistic HDC descriptors for each entry in self.Ds. 70 | 71 | Returns: 72 | np.ndarray: A two-dimensional array with the shape (len(self.Ds), self.nDims), 73 | containing the holistic HDC descriptors for each entry in self.Ds. 74 | """ 75 | Y = np.zeros((len(self.Ds), self.nDims), 'float32') 76 | 77 | for i in range(len(self.Ds)): 78 | # load i-th descriptor 79 | D = self.Ds[i] 80 | 81 | # standardize local descriptors 82 | D['descriptors'] = D['descriptors'] @ self.Proj 83 | D['descriptors'] = self.__STD(D['descriptors']) 84 | 85 | # compute holistic HDC-descriptor 86 | Y[i, :] = self.__bundleLocalDescriptorsIndividually(D) 87 | 88 | return Y 89 | 90 | 91 | def __bundleLocalDescriptorsIndividually(self, D: Dict[str, Union[np.ndarray, int]]) -> np.ndarray: 92 | """ 93 | Binds each local descriptor to its pose and bundles them to compute the holistic 94 | HDC descriptor for a given input dictionary D. 95 | 96 | Args: 97 | D (Dict[str, Union[np.ndarray, int]]): A dictionary containing the following keys: 98 | - 'descriptors': A two-dimensional array containing the local descriptors. 99 | - 'keypoints': A two-dimensional array containing the keypoints. 100 | - 'imheight': An integer representing the image height. 101 | - 'imwidth': An integer representing the image width. 102 | 103 | Returns: 104 | np.ndarray: A one-dimensional array of length self.nDims, containing the 105 | holistic HDC descriptor for the input dictionary D. 106 | """ 107 | desc = D['descriptors'] 108 | keypoints = D['keypoints'] 109 | h = D['imheight'] 110 | w = D['imwidth'] 111 | nDims = self.nDims 112 | 113 | # encode poses 114 | PV = self.__encodePosesHDCconcatMultiAttractor( 115 | keypoints, h, w, nDims) 116 | 117 | # bind each descriptor to its pose and bundle 118 | H = np.sum(desc * PV, 0) 119 | return H 120 | 121 | 122 | def __encodePosesHDCconcatMultiAttractor(self, P: np.ndarray, h: int, w: int, nDims: int) -> np.ndarray: 123 | """ 124 | Encodes the poses of keypoints using the HDC multi-attractor approach. 125 | 126 | Args: 127 | P (np.ndarray): A two-dimensional array containing the keypoints. 128 | h (int): The image height. 129 | w (int): The image width. 130 | nDims (int): The number of dimensions for the HDC vectors. 131 | 132 | Returns: 133 | np.ndarray: A two-dimensional array with the shape (P.shape[0], nDims), 134 | containing the encoded HDC poses for the keypoints. 135 | """ 136 | # relative poses for keypoints 137 | xr = P[:, 1] / w 138 | yr = P[:, 0] / h 139 | 140 | PV = np.zeros((P.shape[0], nDims), 'float32') 141 | for i in range(xr.size): 142 | # find attractors and split index 143 | Xidx1, Xidx2, XsplitIdx = self.__findAttractorsAndSplitIdx( 144 | xr[i], self.posX, nDims) 145 | Yidx1, Yidx2, YsplitIdx = self.__findAttractorsAndSplitIdx( 146 | yr[i], self.posY, nDims) 147 | 148 | # apply 149 | xVec = np.concatenate( 150 | (self.X[Xidx1, :XsplitIdx], self.X[Xidx2, XsplitIdx:])) 151 | yVec = np.concatenate( 152 | (self.Y[Yidx1, :YsplitIdx], self.Y[Yidx2, YsplitIdx:])) 153 | 154 | # combine 155 | PV[i, :] = xVec * yVec 156 | 157 | return PV 158 | 159 | 160 | def __findAttractorsAndSplitIdx(self, p: float, pos: np.ndarray, nDims: int) -> Tuple[int, int, int]: 161 | """ 162 | Finds the two closest attractors to the given position and computes the 163 | split index for combining the HDC vectors. 164 | 165 | Args: 166 | p (float): The position of the point in question. 167 | pos (np.ndarray): A one-dimensional array containing the attractor positions. 168 | nDims (int): The number of dimensions for the HDC vectors. 169 | 170 | Returns: 171 | Tuple[int, int, int]: A tuple containing: 172 | - The index of the first closest attractor. 173 | - The index of the second closest attractor. 174 | - The split index for combining the HDC vectors. 175 | """ 176 | # find closest vectors 177 | idx = np.argpartition(abs(pos-p), 2)[:2] 178 | idx.sort() 179 | idx1, idx2 = idx 180 | 181 | # find weighting of both 182 | d1 = abs(pos[idx1]-p) 183 | d2 = abs(pos[idx2]-p) 184 | w = d2 / (d1 + d2) 185 | 186 | # compute index from weighting 187 | splitIdx = round(w*nDims) 188 | 189 | # return indices 190 | return idx1, idx2, splitIdx 191 | 192 | 193 | def __STD(self, D: np.ndarray) -> np.ndarray: 194 | """ 195 | Standardizes the input descriptors by subtracting the mean and dividing by the standard deviation. 196 | 197 | Args: 198 | D (np.ndarray): A two-dimensional array containing the descriptors to be standardized. 199 | 200 | Returns: 201 | np.ndarray: A two-dimensional array containing the standardized descriptors. 202 | """ 203 | 204 | mu = D.mean(0) 205 | sig = D.std(0) 206 | 207 | D = (D-mu) / sig 208 | 209 | return D 210 | -------------------------------------------------------------------------------- /feature_extraction/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/feature_extraction/__init__.py -------------------------------------------------------------------------------- /feature_extraction/feature_extractor.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | 19 | from abc import ABC, abstractmethod 20 | import numpy as np 21 | from typing import List 22 | 23 | 24 | class FeatureExtractor(ABC): 25 | 26 | @abstractmethod 27 | def compute_features(self, imgs: List[np.ndarray]) -> np.ndarray: 28 | pass 29 | -------------------------------------------------------------------------------- /feature_extraction/feature_extractor_cosplace.py: -------------------------------------------------------------------------------- 1 | 2 | import torch 3 | from torch.utils.data import DataLoader 4 | import torchvision.transforms as transforms 5 | import torch.utils.data as data 6 | 7 | from typing import List 8 | import numpy as np 9 | from tqdm.auto import tqdm 10 | 11 | 12 | class ImageDataset(data.Dataset): 13 | def __init__(self, imgs): 14 | super().__init__() 15 | self.mytransform = self.input_transform() 16 | self.images = imgs 17 | 18 | def __getitem__(self, index): 19 | img = self.images[index] 20 | img = self.mytransform(img) 21 | 22 | return img, index 23 | 24 | def __len__(self): 25 | return len(self.images) 26 | 27 | @staticmethod 28 | def input_transform(): 29 | return transforms.Compose([ 30 | transforms.ToPILImage(), 31 | transforms.Resize(480), 32 | transforms.ToTensor(), 33 | transforms.Normalize(mean=[0.485, 0.456, 0.406], 34 | std=[0.229, 0.224, 0.225]), 35 | ]) 36 | 37 | 38 | class CosPlaceFeatureExtractor(torch.nn.Module): 39 | def __init__(self): 40 | super().__init__() 41 | 42 | if torch.cuda.is_available(): 43 | print('Using GPU') 44 | self.device = torch.device("cuda") 45 | elif torch.backends.mps.is_available() and torch.backends.mps.is_built(): 46 | print('Using MPS') 47 | self.device = torch.device("mps") 48 | else: 49 | print('Using CPU') 50 | self.device = torch.device("cpu") 51 | self.model = torch.hub.load("gmberton/cosplace", "get_trained_model", 52 | backbone="ResNet50", fc_output_dim=2048) 53 | self.dim = 2048 54 | self.model = self.model.to(self.device) 55 | 56 | def compute_features(self, imgs: List[np.ndarray]) -> np.ndarray: 57 | img_set = ImageDataset(imgs) 58 | test_data_loader = DataLoader(dataset=img_set, num_workers=4, 59 | batch_size=4, shuffle=False, 60 | pin_memory=torch.cuda.is_available()) 61 | self.model.eval() 62 | with torch.no_grad(): 63 | global_feats = np.empty((len(img_set), self.dim), dtype=np.float32) 64 | for input_data, indices in tqdm(test_data_loader): 65 | indices_np = indices.numpy() 66 | input_data = input_data.to(self.device) 67 | image_encoding = self.model(input_data) 68 | global_feats[indices_np, :] = image_encoding.cpu().numpy() 69 | return global_feats 70 | 71 | 72 | -------------------------------------------------------------------------------- /feature_extraction/feature_extractor_eigenplaces.py: -------------------------------------------------------------------------------- 1 | 2 | import torch 3 | from torch.utils.data import DataLoader 4 | import torchvision.transforms as transforms 5 | import torch.utils.data as data 6 | 7 | from typing import List 8 | import numpy as np 9 | from tqdm.auto import tqdm 10 | 11 | 12 | class ImageDataset(data.Dataset): 13 | def __init__(self, imgs): 14 | super().__init__() 15 | self.mytransform = self.input_transform() 16 | self.images = imgs 17 | 18 | def __getitem__(self, index): 19 | img = self.images[index] 20 | img = self.mytransform(img) 21 | 22 | return img, index 23 | 24 | def __len__(self): 25 | return len(self.images) 26 | 27 | @staticmethod 28 | def input_transform(): 29 | return transforms.Compose([ 30 | transforms.ToPILImage(), 31 | transforms.Resize(480), 32 | transforms.ToTensor(), 33 | transforms.Normalize(mean=[0.485, 0.456, 0.406], 34 | std=[0.229, 0.224, 0.225]), 35 | ]) 36 | 37 | 38 | class EigenPlacesFeatureExtractor(torch.nn.Module): 39 | def __init__(self): 40 | super().__init__() 41 | 42 | if torch.cuda.is_available(): 43 | print('Using GPU') 44 | self.device = torch.device("cuda") 45 | elif torch.backends.mps.is_available() and torch.backends.mps.is_built(): 46 | print('Using MPS') 47 | self.device = torch.device("mps") 48 | else: 49 | print('Using CPU') 50 | self.device = torch.device("cpu") 51 | self.model = torch.hub.load("gmberton/eigenplaces", "get_trained_model", 52 | backbone="ResNet50", fc_output_dim=2048) 53 | self.dim = 2048 54 | self.model = self.model.to(self.device) 55 | 56 | def compute_features(self, imgs: List[np.ndarray]) -> np.ndarray: 57 | img_set = ImageDataset(imgs) 58 | test_data_loader = DataLoader(dataset=img_set, num_workers=4, 59 | batch_size=4, shuffle=False, 60 | pin_memory=torch.cuda.is_available()) 61 | self.model.eval() 62 | with torch.no_grad(): 63 | global_feats = np.empty((len(img_set), self.dim), dtype=np.float32) 64 | for input_data, indices in tqdm(test_data_loader): 65 | indices_np = indices.numpy() 66 | input_data = input_data.to(self.device) 67 | image_encoding = self.model(input_data) 68 | global_feats[indices_np, :] = image_encoding.cpu().numpy() 69 | return global_feats 70 | 71 | 72 | -------------------------------------------------------------------------------- /feature_extraction/feature_extractor_holistic.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | 19 | import numpy as np 20 | from typing import List 21 | 22 | from .feature_extractor import FeatureExtractor 23 | 24 | 25 | class AlexNetConv3Extractor(FeatureExtractor): 26 | def __init__(self, nDims: int = 4096): 27 | import torch 28 | from torchvision import transforms 29 | 30 | self.nDims = nDims 31 | # load alexnet 32 | self.model = torch.hub.load('pytorch/vision:v0.10.0', 'alexnet', pretrained=True) 33 | 34 | # select conv3 35 | self.model = self.model.features[:7] 36 | 37 | # preprocess images 38 | self.preprocess = transforms.Compose([ 39 | transforms.ToPILImage(), 40 | transforms.Resize([224, 224]), 41 | # transforms.CenterCrop(224), 42 | transforms.ToTensor(), 43 | transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), 44 | ]) 45 | 46 | if torch.cuda.is_available(): 47 | print('Using GPU') 48 | self.device = torch.device("cuda") 49 | elif torch.backends.mps.is_available() and torch.backends.mps.is_built(): 50 | print('Using MPS') 51 | self.device = torch.device("mps") 52 | else: 53 | print('Using CPU') 54 | self.device = torch.device("cpu") 55 | 56 | self.model.to(self.device) 57 | 58 | 59 | def compute_features(self, imgs: List[np.ndarray]) -> List[np.ndarray]: 60 | import torch 61 | 62 | imgs_torch = [self.preprocess(img) for img in imgs] 63 | imgs_torch = torch.stack(imgs_torch, dim=0) 64 | 65 | imgs_torch = imgs_torch.to(self.device) 66 | 67 | with torch.no_grad(): 68 | output = self.model(imgs_torch) 69 | 70 | output = output.to('cpu').numpy() 71 | Ds = output.reshape([len(imgs), -1]) 72 | 73 | rng = np.random.default_rng(seed=0) 74 | Proj = rng.standard_normal([Ds.shape[1], self.nDims], 'float32') 75 | Proj = Proj / np.linalg.norm(Proj , axis=1, keepdims=True) 76 | 77 | Ds = Ds @ Proj 78 | 79 | return Ds 80 | 81 | 82 | class HDCDELF(FeatureExtractor): 83 | def __init__(self): 84 | from .feature_extractor_local import DELF 85 | 86 | self.DELF = DELF() # local DELF descriptor 87 | 88 | def compute_features(self, imgs: List[np.ndarray]) -> np.ndarray: 89 | from feature_aggregation.hdc import HDC 90 | 91 | D_local = self.DELF.compute_features(imgs) 92 | D_holistic = HDC(D_local).compute_holistic() 93 | 94 | return D_holistic 95 | 96 | 97 | # sum of absolute differences (SAD) [Milford and Wyeth (2012). "SeqSLAM: Visual 98 | # Route-Based Navigation for Sunny Summer Days and Stormy Winter Nights". ICRA.] 99 | class SAD(FeatureExtractor): 100 | def __init__(self, nPixels: int = 2048, patchLength: int = 8): 101 | self.nPixels = nPixels # number pixels in downsampled image 102 | self.patchLength = patchLength # side length of patches for patch normalization 103 | 104 | self.imshapeDownsampled = None 105 | 106 | def compute_features(self, imgs: List[np.ndarray]) -> np.ndarray: 107 | 108 | # determine new image shape to obtain roughly self.nPixels pixels and 109 | # image dimensions that are a multiple of self.patchLength 110 | if self.imshapeDownsampled is None: 111 | [h,w,_] = np.array(imgs[0].shape) 112 | 113 | k = np.sqrt(self.nPixels / (h * w)) 114 | h = np.ceil(k * h) 115 | h -= np.mod(h, self.patchLength) 116 | w = np.ceil(k * w) 117 | w -= np.mod(w, self.patchLength) 118 | 119 | if np.abs(self.nPixels - h*w) > np.abs(self.nPixels - (h+self.patchLength)*(w+self.patchLength)): 120 | h += self.patchLength 121 | w += self.patchLength 122 | 123 | self.imshapeDownsampled = [int(h), int(w)] 124 | 125 | # grayscale conversion and downsampling 126 | from torchvision import transforms 127 | preprocess = transforms.Compose([ 128 | transforms.ToPILImage(), 129 | transforms.Grayscale(), 130 | transforms.Resize(self.imshapeDownsampled), 131 | ]) 132 | imgs = [np.array(preprocess(img)) for img in imgs] 133 | 134 | # patch normalization 135 | Ds = [self.__patch_normalize(img).flatten() for img in imgs] 136 | Ds = np.array(Ds).astype('float32') 137 | 138 | return Ds 139 | 140 | def __patch_normalize(self, img: np.ndarray) -> np.ndarray: 141 | np.seterr(divide='ignore', invalid='ignore') # ignore potential division by 0 142 | 143 | img = img.astype('float32') 144 | [h,w] = img.shape 145 | for i_h in range(h // self.patchLength): 146 | for i_w in range(w // self.patchLength): 147 | patch = img[i_h*self.patchLength:(i_h+1)*self.patchLength, i_w*self.patchLength:(i_w+1)*self.patchLength] 148 | patch_normalized = 255 * ((patch - patch.min()) / (patch.max() - patch.min())) 149 | patch = patch_normalized.round() 150 | img[i_h*self.patchLength:(i_h+1)*self.patchLength, i_w*self.patchLength:(i_w+1)*self.patchLength] = patch 151 | 152 | np.seterr(divide='warn', invalid='warn') 153 | 154 | return img 155 | -------------------------------------------------------------------------------- /feature_extraction/feature_extractor_local.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | 19 | import numpy as np 20 | from typing import List 21 | from abc import abstractmethod 22 | 23 | from tqdm.auto import tqdm 24 | 25 | from .feature_extractor import FeatureExtractor 26 | 27 | 28 | class LocalFeatureExtractor(FeatureExtractor): 29 | 30 | @abstractmethod 31 | def compute_local_features(self, imgs: List[np.ndarray]) -> List[np.ndarray]: 32 | pass 33 | 34 | 35 | class DELF(LocalFeatureExtractor): 36 | def __init__(self): 37 | import tensorflow_hub as hub 38 | 39 | self.delf = hub.load('https://tfhub.dev/google/delf/1').signatures['default'] 40 | 41 | def compute_features(self, imgs: List[np.ndarray]) -> np.ndarray: 42 | 43 | D_local = self.compute_local_features(imgs) 44 | 45 | return D_local 46 | 47 | def compute_local_features(self, imgs: List[np.ndarray]) -> List[np.ndarray]: 48 | D = [] 49 | for img in tqdm(imgs): 50 | D.append(self.compute_local_delf_descriptor(img)) 51 | 52 | return D 53 | 54 | def compute_local_delf_descriptor(self, img: np.ndarray): 55 | import tensorflow as tf 56 | 57 | im_height = img.shape[0] 58 | im_width = img.shape[1] 59 | img = tf.image.convert_image_dtype(img, tf.float32) 60 | 61 | out = self.delf(image=img, 62 | score_threshold=tf.constant(0.0), 63 | image_scales=tf.constant([1.0]), 64 | max_feature_num=tf.constant(200)) 65 | 66 | return {'descriptors': np.array(out['features']), 67 | 'descriptors_pca': np.array(out['descriptors']), 68 | 'scores': np.array(out['scores']), 69 | 'keypoints': np.array(out['locations']), 70 | 'scales': 1./np.array(out['scales']), 71 | 'imheight': np.array(im_height), 72 | 'imwidth': np.array(im_width) 73 | } 74 | -------------------------------------------------------------------------------- /feature_extraction/feature_extractor_patchnetvlad.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch.utils.data import DataLoader 3 | import torchvision.transforms as transforms 4 | import torch.utils.data as data 5 | import subprocess 6 | 7 | from PIL import Image 8 | 9 | from os.path import join, isfile 10 | from typing import List 11 | import numpy as np 12 | from tqdm.auto import tqdm 13 | 14 | from patchnetvlad.models.models_generic import get_backend, get_model, get_pca_encoding 15 | from patchnetvlad.tools import PATCHNETVLAD_ROOT_DIR 16 | 17 | from .feature_extractor import FeatureExtractor 18 | 19 | 20 | class ImageDataset(data.Dataset): 21 | def __init__(self, imgs): 22 | super().__init__() 23 | self.mytransform = self.input_transform() 24 | self.images = imgs 25 | 26 | def __getitem__(self, index): 27 | # img = Image.open(self.images[index]) 28 | # TODO: Check if the channel order is correct 29 | img = self.images[index] 30 | img = self.mytransform(img) 31 | 32 | return img, index 33 | 34 | def __len__(self): 35 | return len(self.images) 36 | 37 | @staticmethod 38 | def input_transform(resize=(480, 640)): 39 | if resize[0] > 0 and resize[1] > 0: 40 | return transforms.Compose([ 41 | transforms.ToPILImage(), 42 | transforms.Resize(resize), 43 | transforms.ToTensor(), 44 | transforms.Normalize(mean=[0.485, 0.456, 0.406], 45 | std=[0.229, 0.224, 0.225]), 46 | ]) 47 | else: 48 | return transforms.Compose([ 49 | transforms.ToPILImage(), 50 | transforms.ToTensor(), 51 | transforms.Normalize(mean=[0.485, 0.456, 0.406], 52 | std=[0.229, 0.224, 0.225]), 53 | ]) 54 | 55 | class PatchNetVLADFeatureExtractor(FeatureExtractor): 56 | def __init__(self, config): 57 | self.config = config 58 | 59 | if torch.cuda.is_available(): 60 | print('Using GPU') 61 | self.device = torch.device("cuda") 62 | elif torch.backends.mps.is_available() and torch.backends.mps.is_built(): 63 | print('Using MPS') 64 | self.device = torch.device("mps") 65 | else: 66 | print('Using CPU') 67 | self.device = torch.device("cpu") 68 | 69 | encoder_dim, encoder = get_backend() 70 | 71 | if self.config['global_params']['num_pcs'] != '0': 72 | resume_ckpt = self.config['global_params']['resumePath'] + self.config['global_params']['num_pcs'] + '.pth.tar' 73 | else: 74 | resume_ckpt = self.config['global_params']['resumePath'] + '.pth.tar' 75 | 76 | if not isfile(resume_ckpt): 77 | resume_ckpt = join(PATCHNETVLAD_ROOT_DIR, resume_ckpt) 78 | if not isfile(resume_ckpt): 79 | print('Downloading Patch-NetVLAD models, this might take a while ...') 80 | subprocess.run(["patchnetvlad-download-models"]) 81 | 82 | 83 | if isfile(resume_ckpt): 84 | print("=> loading checkpoint '{}'".format(resume_ckpt)) 85 | checkpoint = torch.load(resume_ckpt, map_location=lambda storage, loc: storage) 86 | if self.config['global_params']['num_pcs'] != '0': 87 | assert checkpoint['state_dict']['WPCA.0.bias'].shape[0] == int(self.config['global_params']['num_pcs']) 88 | self.config['global_params']['num_clusters'] = str(checkpoint['state_dict']['pool.centroids'].shape[0]) 89 | 90 | if self.config['global_params']['num_pcs'] != '0': 91 | use_pca = True 92 | else: 93 | use_pca = False 94 | self.model = get_model(encoder, encoder_dim, self.config['global_params'], append_pca_layer=use_pca) 95 | self.model.load_state_dict(checkpoint['state_dict']) 96 | 97 | if int(self.config['global_params']['nGPU']) > 1 and torch.cuda.device_count() > 1: 98 | self.model.encoder = torch.nn.DataParallel(self.model.encoder) 99 | self.model.pool = torch.nn.DataParallel(self.model.pool) 100 | 101 | self.model = self.model.to(self.device) 102 | print(f"=> loaded checkpoint '{resume_ckpt}'") 103 | else: 104 | raise FileNotFoundError(f"=> no checkpoint found at '{resume_ckpt}'") 105 | 106 | if self.config['global_params']['pooling'].lower() == 'patchnetvlad': 107 | self.num_patches = self.get_num_patches() 108 | else: 109 | self.num_patches = None 110 | 111 | 112 | def get_num_patches(self): 113 | H = int(int(self.config['feature_match']['imageresizeH']) / 16) # 16 is the vgg scaling from image space to feature space (conv5) 114 | W = int(int(self.config['feature_match']['imageresizeW']) / 16) 115 | padding_size = [0, 0] 116 | patch_sizes = [int(s) for s in self.config['global_params']['patch_sizes'].split(",")] 117 | patch_size = (int(patch_sizes[0]), int(patch_sizes[0])) 118 | strides = [int(s) for s in self.config['global_params']['strides'].split(",")] 119 | stride = (int(strides[0]), int(strides[0])) 120 | 121 | Hout = int((H + (2 * padding_size[0]) - patch_size[0]) / stride[0] + 1) 122 | Wout = int((W + (2 * padding_size[1]) - patch_size[1]) / stride[1] + 1) 123 | 124 | num_regions = Hout * Wout 125 | return num_regions 126 | 127 | 128 | def compute_features(self, imgs: List[np.ndarray]) -> np.ndarray: 129 | pool_size = int(self.config['global_params']['num_pcs']) 130 | 131 | img_set = ImageDataset(imgs) 132 | test_data_loader = DataLoader(dataset=img_set, num_workers=int(self.config['global_params']['threads']), 133 | batch_size=int(self.config['feature_extract']['cacheBatchSize']), 134 | shuffle=False, pin_memory=torch.cuda.is_available()) 135 | 136 | self.model.eval() 137 | with torch.no_grad(): 138 | global_feats = np.empty((len(img_set), pool_size), dtype=np.float32) 139 | if self.config['global_params']['pooling'].lower() == 'patchnetvlad': 140 | patch_feats = np.empty((len(img_set), pool_size, self.num_patches), dtype=np.float32) 141 | 142 | for iteration, (input_data, indices) in \ 143 | enumerate(tqdm(test_data_loader), 1): 144 | indices_np = indices.detach().numpy() 145 | input_data = input_data.to(self.device) 146 | image_encoding = self.model.encoder(input_data) 147 | 148 | if self.config['global_params']['pooling'].lower() == 'patchnetvlad': 149 | vlad_local, vlad_global = self.model.pool(image_encoding) 150 | 151 | vlad_global_pca = get_pca_encoding(self.model, vlad_global) 152 | global_feats[indices_np, :] = vlad_global_pca.detach().cpu().numpy() 153 | 154 | for this_local in vlad_local: 155 | patch_feats_batch = np.empty((this_local.size(0), pool_size, this_local.size(2)), 156 | dtype=np.float32) 157 | grid = np.indices((1, this_local.size(0))) 158 | this_local_pca = get_pca_encoding(self.model, this_local.permute(2, 0, 1).reshape(-1, this_local.size(1))).\ 159 | reshape(this_local.size(2), this_local.size(0), pool_size).permute(1, 2, 0) 160 | patch_feats_batch[grid, :, :] = this_local_pca.detach().cpu().numpy() 161 | for i, val in enumerate(indices_np): 162 | patch_feats[val] = patch_feats_batch[i] 163 | else: 164 | vlad_global = self.model.pool(image_encoding) 165 | vlad_global_pca = get_pca_encoding(self.model, vlad_global) 166 | global_feats[indices_np, :] = vlad_global_pca.detach().cpu().numpy() 167 | 168 | if self.config['global_params']['pooling'].lower() == 'patchnetvlad': 169 | return global_feats, patch_feats 170 | else: 171 | return global_feats 172 | 173 | 174 | def local_matcher_from_numpy_single_scale(self, input_query_local_features_prefix, input_index_local_features_prefix): 175 | from patchnetvlad.models.local_matcher import normalise_func, calc_keypoint_centers_from_patches 176 | from patchnetvlad.tools.patch_matcher import PatchMatcher 177 | 178 | patch_sizes = [int(s) for s in self.config['global_params']['patch_sizes'].split(",")] 179 | assert(len(patch_sizes) == 1) 180 | strides = [int(s) for s in self.config['global_params']['strides'].split(",")] 181 | patch_weights = np.array(self.config['feature_match']['patchWeights2Use'].split(",")).astype(float) 182 | 183 | all_keypoints = [] 184 | all_indices = [] 185 | 186 | for patch_size, stride in zip(patch_sizes, strides): 187 | # we currently only provide support for square patches, but this can be easily modified for future works 188 | keypoints, indices = calc_keypoint_centers_from_patches(self.config['feature_match'], patch_size, patch_size, stride, stride) 189 | all_keypoints.append(keypoints) 190 | all_indices.append(indices) 191 | 192 | raw_diffs = [] 193 | 194 | matcher = PatchMatcher(self.config['feature_match']['matcher'], patch_sizes, strides, all_keypoints, all_indices) 195 | 196 | for q_idx in tqdm(range(len(input_query_local_features_prefix)), leave=False, desc='Patch compare pred'): 197 | diffs = np.zeros((len(input_index_local_features_prefix), len(patch_sizes))) 198 | qfeat = [torch.transpose(torch.tensor(input_query_local_features_prefix[q_idx], device=self.device), 0, 1)] 199 | 200 | for candidate in range(len(input_index_local_features_prefix)): 201 | dbfeat = [torch.tensor(input_index_local_features_prefix[candidate], device=self.device)] 202 | diffs[candidate, :], _, _ = matcher.match(qfeat, dbfeat) 203 | 204 | diffs = normalise_func(diffs, len(patch_sizes), patch_weights) 205 | raw_diffs.append(diffs) 206 | 207 | return np.array(raw_diffs).T * -1 208 | -------------------------------------------------------------------------------- /matching/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/matching/__init__.py -------------------------------------------------------------------------------- /matching/matching.py: -------------------------------------------------------------------------------- 1 | # ===================================================================== 2 | # Copyright (C) 2023 Stefan Schubert, stefan.schubert@etit.tu-chemnitz.de 3 | # 4 | # This program is free software: you can redistribute it and/or modify 5 | # it under the terms of the GNU General Public License as published by 6 | # the Free Software Foundation, either version 3 of the License, or 7 | # (at your option) any later version. 8 | # 9 | # This program is distributed in the hope that it will be useful, 10 | # but WITHOUT ANY WARRANTY; without even the implied warranty of 11 | # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 | # GNU General Public License for more details. 13 | # 14 | # You should have received a copy of the GNU General Public License 15 | # along with this program. If not, see . 16 | # ===================================================================== 17 | # 18 | import numpy as np 19 | from scipy.stats import norm 20 | from typing import Union 21 | 22 | 23 | def best_match_per_query(S: np.ndarray) -> np.ndarray: 24 | """ 25 | Finds the best match per query from a similarity matrix S. 26 | 27 | Args: 28 | S (np.ndarray): A two-dimensional similarity matrix with continuous values. 29 | Higher values indicate higher similarity. 30 | 31 | Returns: 32 | np.ndarray: A two-dimensional boolean matrix with the same shape as S, 33 | where the best match per query is marked as True. 34 | """ 35 | i = np.argmax(S, axis=0) 36 | j = np.int64(range(len(i))) 37 | 38 | M = np.zeros_like(S, dtype='bool') 39 | M[i, j] = True 40 | 41 | return M 42 | 43 | 44 | def thresholding(S: np.ndarray, thresh: Union[str, float]) -> np.ndarray: 45 | """ 46 | Applies thresholding on a similarity matrix S based on a given threshold or 47 | an automatic thresholding method. 48 | 49 | The automatic thresholding is based on Eq. 2-4 in Schubert, S., Neubert, P. & 50 | Protzel, P. (2021). Beyond ANN: Exploiting Structural Knowledge for Efficient 51 | Place Recognition. In Proc. of Intl. Conf. on Robotics and Automation (ICRA). 52 | DOI: 10.1109/ICRA48506.2021.9561006 53 | 54 | Args: 55 | S (np.ndarray): A two-dimensional similarity matrix with continuous values. 56 | Higher values indicate higher similarity. 57 | thresh (Union[str, float]): A threshold value or the string 'auto' to apply 58 | automatic thresholding. 59 | 60 | Returns: 61 | np.ndarray: A two-dimensional boolean matrix with the same shape as S, 62 | where values greater or equal to the threshold are marked as True. 63 | """ 64 | if thresh == 'auto': 65 | mu = np.median(S) 66 | sig = np.median(np.abs(S - mu)) / 0.675 67 | thresh = norm.ppf(1 - 1e-6, loc=mu, scale=sig) 68 | 69 | M = S >= thresh 70 | 71 | return M 72 | -------------------------------------------------------------------------------- /output_images/examples_tp_fp.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/output_images/examples_tp_fp.jpg -------------------------------------------------------------------------------- /output_images/matchings.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/output_images/matchings.jpg -------------------------------------------------------------------------------- /output_images/pr_curve.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/stschubert/VPR_Tutorial/e071309301f7b360449700371a69242ed65cad80/output_images/pr_curve.jpg -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | ipywidgets==8.0.6 2 | matplotlib==3.7.1 3 | numpy==1.24.3 4 | patchnetvlad==0.1.7 5 | Pillow==9.5.0 6 | scipy==1.9.1 7 | setuptools==67.7.2 8 | scikit_image==0.19.3 9 | tensorflow==2.11.1 10 | tensorflow_hub==0.12.0 11 | torch==2.0.0 12 | torchvision==0.15.1a0 13 | tqdm==4.65.0 14 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | import os, sys 2 | from setuptools import setup, find_packages 3 | 4 | 5 | here = os.path.abspath(os.path.dirname(__file__)) 6 | 7 | # Get the long description from the README file 8 | with open(os.path.join(here, 'README.md'), encoding='utf-8') as f: 9 | long_description = f.read() 10 | 11 | 12 | install_require_list = [ 13 | 'numpy', 'matplotlib', 'Pillow', 'scipy', 14 | 'scikit-image', 'tensorflow', 'tensorflow_hub', 15 | 'torch', 'torchvision', 'tqdm'] 16 | 17 | # workaround as opencv-python does not show up in "pip list" within a conda environment 18 | # we do not care as conda recipe has py-opencv requirement anyhow 19 | is_conda = os.path.exists(os.path.join(sys.prefix, 'conda-meta')) 20 | if not is_conda: 21 | install_require_list.append('opencv-python') 22 | 23 | setup(name='vpr_tutorial', 24 | version='1.0.0', 25 | description='Visual Place Recognition: A Tutorial. Code repository supplementing our paper.', 26 | long_description = long_description, 27 | long_description_content_type='text/markdown', 28 | author='Stefan Schubert, Peer Neubert, Sourav Garg, Michael Milford and Tobias Fischer', 29 | author_email='stefan.schubert@etit.tu-chemnitz.de', 30 | url='https://github.com/QVPR/Patch-NetVLAD', 31 | license='GPL-3.0-or-later', 32 | classifiers=[ 33 | # 3 - Alpha 34 | # 4 - Beta 35 | # 5 - Production/Stable 36 | 'Development Status :: 4 - Beta', 37 | 38 | # Indicate who your project is intended for 39 | 'Intended Audience :: Developers', 40 | # Pick your license as you wish (should match "license" above) 41 | 'License :: OSI Approved :: GNU General Public License v3 or later (GPLv3+)', 42 | 43 | # Specify the Python versions you support here. In particular, ensure 44 | # that you indicate whether you support Python 2, Python 3 or both. 45 | 'Programming Language :: Python :: 3.8', 46 | 'Programming Language :: Python :: 3.9', 47 | 'Programming Language :: Python :: 3.10', 48 | 'Programming Language :: Python :: 3.11', 49 | ], 50 | python_requires='>=3.8', 51 | install_requires=install_require_list, 52 | packages=find_packages(), 53 | keywords=[ 54 | 'python', 'place recognition', 'image retrieval', 'computer vision', 'robotics' 55 | ], 56 | scripts=['demo.py'], 57 | entry_points={ 58 | 'console_scripts': ['vpr-tutorial-demo=demo:main',], 59 | }, 60 | # package_data={'': ['configs/*.ini', 'dataset_gt_files/*.npz', 'example_images/*', 61 | # 'output_features/.hidden', 'pretrained_models/.hidden', 'results/.hidden', 62 | # 'dataset_imagenames/*.txt']} 63 | ) 64 | --------------------------------------------------------------------------------