DeepWindows

Dataset and PyTorch Implementation of DeepWindows: Windows Instance Segmentation through an Improved Mask R-CNN Using Spatial Attention and Relation Modules.

Dataset
Citation
Requirements
Usage

Dataset

The dataset can be downloaded from Google Drive or 百度网盘 (提取码：srfo)

You need to change the dataset path in train_net.py to your own path.

Citation

If you use our code or dataset, please use the following BibTeX entry.

@Article{sun2022deepwindows,
  author =       {Sun, Yanwei and Malihi, Shirin and Li, Hao and Maboudi, Mehdi},
  title =        {DeepWindows: Windows Instance Segmentation through an Improved Mask R-CNN Using Spatial Attention and Relation Modules},
  journal =      {ISPRS International Journal of Geo-Information},
  volume =       {11},
  year =         {2022},
  number =       {3},
  article-number = {162},
  url =          {https://www.mdpi.com/2220-9964/11/3/162}
  issn =         {2220-9964},
  doi =          {10.3390/ijgi11030162}
}

Requirements

Linux(tested on Ubuntu 20.04)
Python 3.7
PyTorch
detectron2
OpenCV (for visualization)

Usage

Train

to train Mask R-CNN:

./train_net.py \
  --config-file ./configs/mask_rcnn_R_50_FPN_1x.yaml \

to train deepwindows network:

./train_net.py \
  --config-file ./configs/CASARPN_RM_R_50_FPN_1x.yaml \

Evaluation

to calculate average precision:

./train_net.py \
  --config-file ./configs/CASARPN_RM_R_50_FPN_1x.yaml \
  --eval-only MODEL.WEIGHTS /path/to/checkpoint_file

to calculate pixel accuracy:

./calcPixelAccuracy.py \
--input /JSON file produced by the model
--dataset /name of the dataset

Predict

./predict_results.py \
  --config-file ./configs/CASARPN_RM_R_50_FPN_1x.yaml
  --input /path/to/input/images
  --output /path/to/output
  --opts
  MODEL.WEIGHTS /path/to/checkpoint_file

SunYW0108 / DeepWindows

readme