├── track3 └── README.md ├── track2 └── README.md ├── track1 └── README.md ├── track6 └── README.md ├── track5 └── README.md ├── track4 └── README.md ├── data └── README.md └── README.md /track3/README.md: -------------------------------------------------------------------------------- 1 | # Track 3: Automatic Bridge Detection in Optical Satellite Images 2 | 3 | In Track 3 the goal is to locate the bridge in the largescale optical satellite images. The coordinates of the bridge are horizontal bounding box. 4 | 5 | ## Baseline Algorithm 6 | Two-stage object detection method Faster RCNN based on ResNet-50 is developed for object detection tracks. 7 | 8 | ### Introduction 9 | It is modified from [mmdetection](https://github.com/open-mmlab/mmdetection). The master branch works with **PyTorch 1.1** or higher. 10 | 11 | - Linux 12 | - Python 3.5+ ([Say goodbye to Python2](https://python3statement.org/)) 13 | - PyTorch 1.1 14 | - CUDA 9.0+ 15 | - NCCL 2+ 16 | - GCC 4.9+ 17 | - [mmcv](https://github.com/open-mmlab/mmcv) 18 | 19 | ### Code 20 | The code will be available soon. 21 | 22 | ### Result 23 | 24 | | Method | Result (mAP) | 25 | | ----------- | ----------- | 26 | | Faster RCNN | 59.52 | -------------------------------------------------------------------------------- /track2/README.md: -------------------------------------------------------------------------------- 1 | # Track 2: Ship Detection in SAR Images 2 | 3 | Track 2 is for the detection of ships in SAR images, whose purpose is to locate the ship in SAR images. In each image, the coordinates of ships are described in a predefined format. The horizontal bounding boxes of ships are provided in XML files. 4 | 5 | ## Baseline Algorithm 6 | Two-stage object detection method Faster RCNN based on ResNet-50 is developed for object detection tracks. 7 | 8 | ### Introduction 9 | It is modified from [mmdetection](https://github.com/open-mmlab/mmdetection). The master branch works with **PyTorch 1.1** or higher. 10 | 11 | - Linux 12 | - Python 3.5+ ([Say goodbye to Python2](https://python3statement.org/)) 13 | - PyTorch 1.1 14 | - CUDA 9.0+ 15 | - NCCL 2+ 16 | - GCC 4.9+ 17 | - [mmcv](https://github.com/open-mmlab/mmcv) 18 | 19 | ### Code 20 | The code will be available soon. 21 | 22 | ### Result 23 | 24 | | Method | Result (mAP) | 25 | | ----------- | ----------- | 26 | | Faster RCNN | 36.13 | -------------------------------------------------------------------------------- /track1/README.md: -------------------------------------------------------------------------------- 1 | # Track 1: Airplane Detection and Recognition in Optical Images 2 | 3 | Track 1 is for the detection and recognition of airplanes in optical remote sensing images. For each image in the dataset, there is an XML file with the same name for describing annotation information. Each airplane instance in images is annotated by the corresponding category information and location with an oriented bounding box. 4 | 5 | ## Baseline Algorithm 6 | Two-stage object detection method Faster RCNN based on ResNet-50 is developed for object detection tracks. For Track 1, add an angle information regression to realize rotated boxes regression. 7 | 8 | ### Introduction 9 | It is modified from [mmdetection](https://github.com/open-mmlab/mmdetection). The master branch works with **PyTorch 1.1** or higher. 10 | 11 | - Linux 12 | - Python 3.5+ ([Say goodbye to Python2](https://python3statement.org/)) 13 | - PyTorch 1.1 14 | - CUDA 9.0+ 15 | - NCCL 2+ 16 | - GCC 4.9+ 17 | - [mmcv](https://github.com/open-mmlab/mmcv) 18 | 19 | ### Code 20 | The code will be available soon. 21 | 22 | ### Result 23 | 24 | | Method | Result (mAP) | 25 | | ----------- | ----------- | 26 | | Faster RCNN | 48.42 | -------------------------------------------------------------------------------- /track6/README.md: -------------------------------------------------------------------------------- 1 | # Track 6: Semantic Segmentation in Fully Polarimetric SAR Images 2 | 3 | Track 6 is a semantic segmentation track for SAR images. Its goal is to classify the features in SAR satellite images with pixel level. 4 | 5 | ## Baseline Algorithm 6 | For semantic segmentation tracks, using PSPNet based on ResNet-50 as baseline solution. 7 | 8 | ### Introduction 9 | PSPNet baseline for semantic segmentation tracks in 2020 Gaofen Challenge is written by the deep learning framework of PyTorch. 10 | 11 | - Linux 12 | - Python 3.5+ 13 | - PyTorch 1.0+ 14 | - CUDA 10.0+ 15 | - NCCL 2+ 16 | - GCC 4.9+ 17 | 18 | 19 | 20 | ### Data format 21 | 22 | * Convert the data format to the same as Vaihingen and modify the ${CONFIG_FILE}. 23 | 24 | Assuming `C` as number of classes in the semantic segmentation dataset, then valid label ids are from `0` to `C-1`. And we tend to set the ignore label as 255 where loss calculation will be ignored and no penalty will be given on the related ground truth regions. If original ground truths ids are not in needed format, you may need to do label id mapping 25 | 26 | * Prepare the `$DATASET$_colors.txt` and `$DATASET$_names.txt` accordingly. Get the training/testing ground truths and lists ready. 27 | 28 | 29 | ### Code 30 | The code will be available soon. 31 | 32 | ### Result 33 | 34 | | Method | Result (FWIoU) | 35 | | ----------- | ----------- | 36 | | PSPNet | 60.68 | -------------------------------------------------------------------------------- /track5/README.md: -------------------------------------------------------------------------------- 1 | # Track 5: Automatic Water-Body Segmentation in Optical Satellite Images 2 | 3 | In Track 5 the purpose is to locate the water-body in the optical satellite images with pixel level. The original optical satellite images and ground truth images were provided for water-body segmentation. 4 | 5 | ## Baseline Algorithm 6 | For semantic segmentation tracks, using PSPNet based on ResNet-50 as baseline solution. 7 | 8 | ### Introduction 9 | PSPNet baseline for semantic segmentation tracks in 2020 Gaofen Challenge is written by the deep learning framework of PyTorch. 10 | 11 | - Linux 12 | - Python 3.5+ 13 | - PyTorch 1.0+ 14 | - CUDA 10.0+ 15 | - NCCL 2+ 16 | - GCC 4.9+ 17 | 18 | 19 | 20 | ### Data format 21 | 22 | * Convert the data format to the same as Vaihingen and modify the ${CONFIG_FILE}. 23 | 24 | Assuming `C` as number of classes in the semantic segmentation dataset, then valid label ids are from `0` to `C-1`. And we tend to set the ignore label as 255 where loss calculation will be ignored and no penalty will be given on the related ground truth regions. If original ground truths ids are not in needed format, you may need to do label id mapping 25 | 26 | * Prepare the `$DATASET$_colors.txt` and `$DATASET$_names.txt` accordingly. Get the training/testing ground truths and lists ready. 27 | 28 | 29 | ### Code 30 | The code will be available soon. 31 | 32 | ### Result 33 | 34 | | Method | Result (FWIoU) | 35 | | ----------- | ----------- | 36 | | PSPNet | 90.07 | -------------------------------------------------------------------------------- /track4/README.md: -------------------------------------------------------------------------------- 1 | # Track 4: Semantic Segmentation in Optical Satellite Images 2 | 3 | Track 4 is for semantic segmentation in optical satellite images. A pair of images are provided for each geographic tile. One is the original optical satellite image, and the other is an image annotated with the ground truth which is the same size as the previous satellite image. In ground truth images, different categories are marked with different RGB values in pixel level. 4 | 5 | ## Baseline Algorithm 6 | For semantic segmentation tracks, using PSPNet based on ResNet-50 as baseline solution. 7 | 8 | ### Introduction 9 | PSPNet baseline for semantic segmentation tracks in 2020 Gaofen Challenge is written by the deep learning framework of PyTorch. 10 | 11 | - Linux 12 | - Python 3.5+ 13 | - PyTorch 1.0+ 14 | - CUDA 10.0+ 15 | - NCCL 2+ 16 | - GCC 4.9+ 17 | 18 | 19 | 20 | ### Data format 21 | 22 | * Convert the data format to the same as Vaihingen and modify the ${CONFIG_FILE}. 23 | 24 | Assuming `C` as number of classes in the semantic segmentation dataset, then valid label ids are from `0` to `C-1`. And we tend to set the ignore label as 255 where loss calculation will be ignored and no penalty will be given on the related ground truth regions. If original ground truths ids are not in needed format, you may need to do label id mapping 25 | 26 | * Prepare the `$DATASET$_colors.txt` and `$DATASET$_names.txt` accordingly. Get the training/testing ground truths and lists ready. 27 | 28 | 29 | ### Code 30 | The code will be available soon. 31 | 32 | ### Result 33 | 34 | | Method | Result (FWIoU) | 35 | | ----------- | ----------- | 36 | | PSPNet | 69.78 | -------------------------------------------------------------------------------- /data/README.md: -------------------------------------------------------------------------------- 1 | # DATA 2 | To satisfy the needs of practical applications and the highresolution earth observation system construction requirements for major national scientific and technological projects, images in Gaofen Challenge are collected from Gaofen-2 satellite and Gaofen-3 satellite. Specifically, we use the Gaofen-2 optical satellite data with 0.8-4m resolution for the Airplane Detection, Bridge Detection, Water-body Segmentation and so on. And the Gaofen-3 SAR data with 1-5m resolution is used for Ship Detection, Semantic Segmentation in Polarimetric SAR and so on. To obtain high-quality data, we invited hundreds of experts taking more than three months to prepare the dataset. Finally, a large-scale and challenging dataset with various categories, multiple sensors and tremendous object instances have been published for Gaofen Challenge. 3 | 4 | ## Data Description: 5 | The Gaofen-2 optical satellite data with 0.8-4m resolution is used for the Airplane Detection, Bridge Detection, Water-body Segmentation and so on. And the Gaofen-3 SAR data with 1-5m resolution is used for Ship Detection, Semantic Segmentation in Polarimetric SAR and so on. 6 | 7 | | Track | DESCRIPTION | 8 | | ----------- | ----------- | 9 | | Track-1 | Data for Track 1 (Airplane Detection and Recognition in Optical Images) were provided by Gaofen-2 satellite. The scenes include the main civil airports in the world, such as Sydney Airport, Beijing Capital International Airport, Shanghai Pudong International Airport, Hongkong Airport, Tokyo International Airport, and so on. It contains 3000 remote sensing images with spatial resolution of 0.8m. Each image is of the size 1000x1000 pixels and contains 10 categories of airplanes (i.e. Boeing 737, Boeing 747, Boeing 777, Boeing 787, Airbus A220, Airbus A321, Airbus A330, Airbus A350, COMAC ARJ21, and other) exhibiting a wide variety of orientations and scales. | 10 | | Track-2 | Data for Track 2 (Ship Detection in SAR Images) were collected from Gaofen-3 satellite. It contains 1000 SAR images with spatial resolution ranging from 1m-5m. Each image is of the size 1000x1000 pixels and contains ships exhibiting a wide variety of orientations and scales. The scenes include the main civil ports in the world, such as Victoria Harbour, Port of Sanya, Incheon Port, and so on. | 11 | | Track-3 | Data for Track 3 (Automatic Bridge Detection in Optical Satellite Images) were provided by Gaofen-2 satellite with the resolution ranging from 1m-4m. Each image contains at least one bridge, covering railway bridges, highway bridges, road-rail bridges, pedestrian bridges, water-carrying bridges, and so on. There are 3000 images with different sizes ranging from 667x667 to 1001x1001 pixels in bridge dataset. | 12 | | Track-4 | Data for Track 4 (Semantic Segmentation in Optical Satellite Images) were provided by Gaofen-2 satellite with 0.8m resolution. Each image was finely annotated with nine categories of ground objects at the pixel level, including road, building, shrub & tree, lawn, land, waterbody, vehicle, impervious ground, and other. There are 1800 images with the size ranging from 512 to 5000 pixels. | 13 | | Track-5 | Data for Track 5 (Automatic Water-Body Segmentation in Optical Satellite Images) were provided by Gaofen-2 satellite with the resolution ranging from 1m-4m, covering rivers and lakes in large-scope. There are 2500 images with the size ranging from 492 to 2000 pixels in water-body dataset. | 14 | | Track-6 | Data for Track 6 (Semantic Segmentation in Fully Polarimetric SAR Images) were provided by Gaofen-3 satellite with 1-3m resolution, containing 4 polarization modes (i.e. HH, VV, HV, and VH). Six categories including water-body, building, industrial area, lawn, land, and others were annotated at the pixel level for each image. There are 1200 images with the size ranging from 512 to 1500 pixels. | 15 | 16 | 17 | ## Image sizes: 18 | 19 | | Track | train set number | preliminary test set number | final test set number | Resolution (pixel) | Size (GB) | 20 | | ------- | -----------------| --------------------------- | ----------------------| ------------------ | --------- | 21 | | Track-1 | 1000 | 1000 | 1000 | 1000x1000 | 5 | 22 | | Track-2 | 300 | 400 | 300 |1000x1000 | 2 | 23 | | Track-3 | 2000 | 1000 | 1500 |512-12000 | 5.3 | 24 | | Track-4 | 500 | 300 | 1000 | 512-5000 | 3.24 | 25 | | Track-5 | 1000 | 500 | 1000 |492-2000 | 1.2 | 26 | | Track-6 | 500 | 300 | 400 | 512-1500 | 2 | 27 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # 2020 Gaofen Challenge data, baselines, and metrics 2 | For more information visit 2020 Gaofen Challenge on Automated High-Resolution Earth Observation Image Interpretation: [Tracks 1-2](http://en.sw.chreos.org/), [Tracks 3-6](http://sw.chreos.org/). 3 | 4 | 5 | ## Cloning this repository 6 | This repository contains large model files used by baseline algorithms. The model files do not need to be downloaded, but are provided as a convenience. Please use [git lfs](https://git-lfs.github.com/) when cloning to have access to these models. If you did not install and initialize git lfs before cloning, you can simply run `git lfs fetch` after locally initializing git lfs. 7 | 8 | ## Challenge Tracks 9 | ### Track 1: Airplane Detection and Recognition in Optical Images 10 | [Track 1](track1) is for the detection and recognition of airplanes in optical remote sensing images. For each image in the dataset, there is an XML file with the same name for describing annotation information. Each airplane instance in images is annotated by the corresponding category information and location with an oriented bounding box. 11 | 12 | ### Track 2: Ship Detection in SAR Images 13 | [Track 2](track2) is for the detection of ships in SAR images, whose purpose is to locate the ship in SAR images. In each image, the coordinates of ships are described in a predefined format. The horizontal bounding boxes of ships are provided in XML files. 14 | 15 | ### Track 3: Automatic Bridge Detection in Optical Satellite Images 16 | In [Track 3](track3) the goal is to locate the bridge in the largescale optical satellite images. The coordinates of the bridge are horizontal bounding box. 17 | 18 | ### Track 4: Semantic Segmentation in Optical Satellite Images 19 | [Track 4](track4) is for semantic segmentation in optical satellite images. A pair of images are provided for each geographic tile. One is the original optical satellite image, and the other is an image annotated with the ground truth which is the same size as the previous satellite image. In ground truth images, different categories are marked with different RGB values in pixel level. 20 | 21 | ### Track 5: Automatic Water-Body Segmentation in Optical Satellite Images 22 | In [Track 5](track5) the purpose is to locate the water-body in the optical satellite images with pixel level. The original optical satellite images and ground truth images were provided for water-body segmentation. 23 | 24 | ### Track 6: Semantic Segmentation in Fully Polarimetric SAR Images 25 | [Track 6](track6) is a semantic segmentation track for SAR 26 | images. Its goal is to classify the features in SAR satellite images with pixel level. 27 | 28 | ## Performance 29 | For the track of object detection (Track 1-3), mean Average Precision (mAP) with the Intersection over Union (IoU) threshold of 0.5 is used to evaluate the results. For a given ground truth and the predicted result, TP, FP, FN are selected according to an IoU threshold of 0.5. Then the precision and recall are calculated. According to Pascal VOC 2012, the AP of each class is calculated based on precision and recall, and then the mAP can be obtained. For the track of semantic segmentation (Track 4- 30 | 6), Frequency Weighted Intersection over Union (FWIoU) is 31 | used as an accuracy evaluation indicator. 32 | 33 | ## Data 34 | Data from Chinese Gaofen satellites were provided for all 35 | six tracks of the 2020 Gaofen Challenge. The data used in the challenge includes multi-scale, multi-view, multi-resolution optical remote sensing images and SAR images, with ground truth geometric and semantic labels. The optical remote sensing images and SAR images were all collected from Gaofen-2 and Gaofen-3 satellites with the resolution ranging from 1-4 meters and 1-5 meters, respectively. The data containing more than 10,000 images were annotated by more than 100 experts over 3 months. 36 | 37 | ## Baseline algorithms 38 | Classic object detection and semantic segmentation networks 39 | are used as baseline solutions of each track separately. 40 | Two-stage object detection method Faster RCNN based on 41 | ResNet-50 is developed for object detection tracks. For Track 1, add an angle information regression to realize rotated boxes regression. For semantic segmentation tracks, using PSPNet based on ResNet-50 as baseline solution. These baselines are available in the Track 1-6 folders referenced above. 42 | 43 | ## Submission requirements 44 | Submissions must match the formats and data types described on the challenge website. Please check your files before submitting. There is no requirement to use a particular language for producing submissions. 45 | 46 | ## Acknowledgements 47 | We are very grateful for the support of the IEEE Geoscience and Remote Sensing Society, especially to Professor Paolo Gamba, Professor Jun Li, and the Image Analysis and Data Fusion Technical Committee for their valuable comments. And many thanks to the International Society for Photogrammetry and Remote Sensing, especially to Professor Charles Toth and Professor Stefan Hinz for their great support. It is also supported by the National Natural Science Foundation of China 61725105 and the National Major Project on China High-resolution Earth Observation System GFZX0404120201. 48 | 49 | ## Reference 50 | For more information on the data of Track 4: 51 | X. Sun, A. Shi, H. Huang, H. Mayer. BAS4Net: Boundary-Aware Semi-Supervised Semantic Segmentation Network for Very High-Resolution Remote Sensing Images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020. DOI: 10.1109/JSTARS.2020.3021098. 52 | 53 | We have constructed the largest remote sensing image fine-grained target recognition data set FAIR1M, and released it to the world. The relevant data set papers have been synchronized and open source. For details, please refer to: FAIR1M: A Benchmark Dataset for Fine-grained Object Recognition in High-Resolution Remote Sensing Imagery[J]. arXiv preprint arXiv:2103.05569, 2021. (Data set paper link address: https://arxiv.org/abs/2103.05569) 54 | 55 | --------------------------------------------------------------------------------