MSNET

This repository contains the code and models for the following WACV'21 paper:

MSNet: A Multilevel Instance Segmentation Network for Natural Disaster Damage Assessment in Aerial Videos

If you find this code useful in your research then please cite

@misc{zhu2020msnet,
    title={MSNet: A Multilevel Instance Segmentation Network for Natural Disaster Damage Assessment in Aerial Videos},
    author={Xiaoyu Zhu and Junwei Liang and Alexander Hauptmann},
    year={2020},
    eprint={2006.16479},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

The ISBDA Dataset

• Download links: Google Drive

• The dataset has 1,030 images with 2,961 annotations. Each annotation includes the damage segmentation, damage bounding boxes, and house bounding boxes.

Code

Training

Replace /usr/local/lib/python3.6/dist-packages/pycocotools/coco.py and /usr/local/lib/python3.6/dist-packages/pycocotools/cocoeval.py with files on /env folder. And run:

$ cd ./examples/msnet/
$ python train.py --config DATA.BASEDIR=data_dir MODE_FPN=True \
  DATA.VAL=('val',)  DATA.TRAIN=('train',)  \
  TRAIN.BASE_LR=1e-3 TRAIN.EVAL_PERIOD=1 TRAIN.LR_SCHEDULE=[3000]  \
  PREPROC.TRAIN_SHORT_EDGE_SIZE=[600,1200] TRAIN.CHECKPOINT_PERIOD=1 DATA.NUM_WORKERS=1 \
  --load  checkpoint_dir\
  --logdir log_dir

Inferencing

$ cd ./examples/msnet/
$ python predict.py \
 --config DATA.BASEDIR=data_dir MODE_FPN=True \
 DATA.VAL=('val',)  DATA.TRAIN=('train',) \
 --load checkpoint_dir --evaluate output_json_file

Google Colab Demo

Follow demo.ipynb to run pre-trained models on custom dataset and visulize the prediction results.

Pre-Trained Model

Download links: Google Drive

License

Our dataset, code, and models are only for ACADEMIC OR NON-PROFIT ORGANIZATION NONCOMMERCIAL RESEARCH USE ONLY.

Acknowledgements

Our MSNet is based on Tensorpack.