README.md

iSAID: A Large-scale Dataset for Instance Segmentation in Aerial Images, CVPR workshops, 2019.

Codes for Data Preparation and Evaluation

1. Environment and dependencies installation

   1. Create the conda environment conda env create -f environment.yml
   2. Activate the current working environment source activate py_isaid
   3. Setup pycocotols for the evalaution server - cd cocoapi/PythonAPI - make - python setup.py install
   4. Setup cityscapesScripts for the evalaution server - cd preprocess/cityscapesScripts - python setup.py install
   5. Setup detectron for the evalaution server - cd preprocess/Detectron - make
   6. Note: opencv version == 3.4.2

2. Data Preparation for Training, Validation and Testing

   1. Please download iSAID dataset that contains image segmentation masks. Also, download original images from DOTA dataset. Make sure that the final dataset must have this structure:

       iSAID
       ├── test
       │   └── images
       │       ├── P0006.png
       │       └── ...
       │       └── P0009.png
       ├── train
       │   └── images
       │       ├── P0002_instance_color_RGB.png
       │       ├── P0002_instance_id_RGB.png
       │       ├── P0002.png
       │       ├── ...
       │       ├── P0010_instance_color_RGB.png
       │       ├── P0010_instance_id_RGB.png
       │       └── P0010.png
       └── val
           └── images
               ├── P0003_instance_color_RGB.png
               ├── P0003_instance_id_RGB.png
               ├── P0003.png
               ├── ...
               ├── P0004_instance_color_RGB.png
               ├── P0004_instance_id_RGB.png
               └── P0004.png
   

   Note that the segmentation masks for the test images are withheld for the evaluation server.

   3. Change the current working directory to preprocess folder. cd preprocess

   4. Create symlink for iSAID dataset as ln -s /path-of-iSAID-dataset ./dataset/

   5. Split training and validation images into patches python split.py --set train,val

   6. Split test images into patches python split.py --set test

   7. Create coco-format json annotation files for train and val split images python preprocess.py --set train,val

      Make sure that the final dataset after preprocesing must have this structure:

   iSAID_patches
   ├── test
   │   └── images
   │       ├── P0006_0_0_800_800.png
   │       └── ...
   │       └── P0009_0_0_800_800.png
   ├── train
   │   └── instance_only_filtered_train.json
   │   └── images
   │       ├── P0002_0_0_800_800_instance_color_RGB.png
   │       ├── P0002_0_0_800_800_instance_id_RGB.png
   │       ├── P0002_0_800_800.png
   │       ├── ...
   │       ├── P0010_0_0_800_800_instance_color_RGB.png
   │       ├── P0010_0_0_800_800_instance_id_RGB.png
   │       └── P0010_0_800_800.png
   └── val
       └── instance_only_filtered_val.json
       └── images
           ├── P0003_0_0_800_800_instance_color_RGB.png
           ├── P0003_0_0_800_800_instance_id_RGB.png
           ├── P0003_0_0_800_800.png
           ├── ...
           ├── P0004_0_0_800_800_instance_color_RGB.png
           ├── P0004_0_0_800_800_instance_id_RGB.png
           └── P0004_0_0_800_800.png
   

3. Method

   1. Run your instance segmentation method on patches and generate json file of predictions

4. Evaluation

   1. Change the current working directory to evaluate folder. cd ../evaluate
   2. Given json of predictions and json of val set ground truth (obtained after preprocess.py), Compute Average Precision python evaluate.py