Rui Yang, Lin Song, Yixiao Ge, Xiu Li

BoxSnake is an end-to-end training technique to achieve effective polygonal instance segmentation using only box annotations. It consists of two loss functions: (1) a point-based unary loss that constrains the bounding box of predicted polygons to achieve coarse-grained segmentation; and (2) a distance-aware pairwise loss that encourages the predicted polygons to fit the object boundaries.

Arxiv Paper | [Video Demo]

• Python=3.8
• Pytorch=1.9.0
• Detectron2: follow Detectron2 installation instructions.
• pip install -r requirements.txt

BoxSnake also uses the deformable attention modules introduced in Deformable-DETR and the differentiable rasterizer introduced in BoundaryFormer. Please build them on your system:

bash scripts/auto_build.sh

or

cd ./modeling/layers/deform_attn
sh ./make.sh
cd ./modeling/layers/diff_ras
python setup.py build install

conda create --name boxsnake python=3.8 -y
conda activate boxsnake

conda install pytorch==1.9.0 torchvision==0.10.0 cudatoolkit=11.1 -c pytorch -c nvidia
python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

git clone https://github.com/Yangr116/BoxSnake.git
cd BoxSnake
pip install -r requirements.txt
bash scripts/auto_build.sh

mask AP is the result on validation set.

We use the COCO dataset and Cityscapes dataset. Please following here to prepare them.

If you would like to use swin transformer backbone, please download swin weights from here and convert them to pkl format:

python tools/convert-pretrained-model-to-d2.py ${your_swin_pretrained.pth} ${yout_swin_pretrained.pkl}

To train on COCO dataset using the R50 backbone at a 1X schedule:

# 8 gpus
python train_net.py --num-gpus 8 --config-file configs/COCO-InstanceSegmentation/BoxSnake_RCNN/boxsnake_R_50_FPN_1x.yaml

You can also run below code:

bash scripts/auto_run.sh $CONFIG  # your config

To inference on COCO validation set using trained weights:

# 8 gpus
python train_net.py --num-gpus 8 --config-file configs/COCO-InstanceSegmentation/BoxSnake_RCNN/boxsnake_R_50_FPN_1x.yaml
 --eval-only MODEL.WEIGHTS ${your/checkpoints/boxsnake_rcnn_R_50_FPN_coco_1x.pth}

Inference on a single image using trained weights:

python demo/demo.py --config-file configs/COCO-InstanceSegmentation/BoxSnake_RCNN/boxsnake_R_50_FPN_1x.yaml --input demo/demo.jpg --output ${/your/visualized/dir} --confidence-threshold 0.5 --opts MODEL.WEIGHTS ${your/checkpoints/boxsnake_rcnn_R_50_FPN_coco_1x.pth}

BoxSnake is inspired by traditional levelset (including boxlevelset) and GVF methods, and you can check below links to learn them:

• https://www.csd.uwo.ca/~yboykov/Presentations/ECCV06_tutorial_partII_dan.pdf
• https://www.youtube.com/watch?v=1ZJ88JyLPZI
• https://agustinus.kristia.de/techblog/2016/11/05/levelset-method
• https://agustinus.kristia.de/techblog/2016/11/20/levelset-segmentation

Some geometric knowledge may help readers to understand the BoxSnake better:

• Points in Polygon (PIP): https://wrfranklin.org/Research/Short_Notes/pnpoly.html#The%20Method
• PIP: https://towardsdatascience.com/is-the-point-inside-the-polygon-574b86472119
• PIP numpy: https://github.com/Dan-Patterson/numpy_geometry/blob/dbc1a00baaf86d8ae437236659a45cfb57f4f35e/arcpro_npg/npg/npg/npg_pip.py
• Distance from a point to a line: http://paulbourke.net/geometry/pointlineplane
• Distance from a point to a polygon: https://stackoverflow.com/questions/10983872/distance-from-a-point-to-a-polygon

If you find BoxSnake helpful, please cite:

@misc{BoxSnake,
      title={BoxSnake: Polygonal Instance Segmentation with Box Supervision}, 
      author={Rui Yang and Lin Song and Yixiao Ge and Xiu Li},
      year={2023},
      eprint={2303.11630},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}