Deeply Shape-guided Cascade for Instance Segmentation

This repo hosts the code for implementing the DSC algorithms for instance segmentation.

Deeply Shape-guided Cascade for Instance Segmentation,
Hao Ding, Siyuan Qiao, Alan Yuille, Wei Shen
In: Proc. IEEE Conference Computer Vision Pattern Recognition(CVPR), 2021
arXiv preprint (arXiv 1911.11263)

Highlights

Imposing shape guidance: DSC improved the original HTC baseline by imposing shape guidance both explicitly and implicitly.
State-of-the-art performance: Our best single model based on ResNeXt-101-64x4d + FPN with deformable convolutions achieves 51.8% in box AP and 45.5% in mask AP on COCO test-dev (without multi-scale testing)
Better improvement for huddled instances: With shape guidance, DSC's improvement is more significant on images with more huddled instances.

Installation

This implementation is based on mmdetection(v2.1.0). Please refer to INSTALL.md GET_STARTED.md for installation and dataset preparation. A list of installed packages information from anaconda is provided in conda_list.txt

Model Zoo

We provide the following trained models on COCO

Model	Backbone	Multi-scale training	Lr schd	box AP	Mask AP	config	Model
F-DSC	R50-FPN	No	1x	44.5	39.5	config	download
DSC	R50-FPN	No	1x	45.0	39.7	config	download
DSC	R50-FPN	No	20e	45.8	40.1	config	download
DSC	R101-FPN	No	20e	46.6	40.7	config	download
DSC	X101-32x4d-FPN	No	20e	48.0	41.9	config	download
DSC	X101-64x4d-DCN-FPN	Yes	20e	51.4	44.9	config	download

Usage

Train with multiple GPUs

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM}

Example: 
./tools/dist_train.sh configs/dsc/fast_dsc_r50_fpn_1x_coco.py  8

Train with single GPU

python tools/train.py ${CONFIG_FILE}

Example:
python tools/train.py configs/dsc/fast_dsc_r50_fpn_1x_coco.py

Testing

# multi-gpu testing
./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} --out  ${OUTPUT_FILE} --eval bbox segm

Example: 
./tools/dist_test.sh configs/dsc/fast_dsc_r50_fpn_1x_coco.py  f_dsc_r50_fpn_1x_coco.pth 8 --out results.pkl --eval bbox segm

# single-gpu testing
python tools/test_ins.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --show --out  ${OUTPUT_FILE} --eval segm

Example: 
python tools/test_ins.py configs/dsc/fast_dsc_r50_fpn_1x_coco.py  f_dsc_r50_fpn_1x_coco.pth --out  results.pkl --eval bbox segm

Citations

Please consider citing our papers in your publications if this repo helps you.

@inproceedings{ding2021dsc,
  title     =  {Deeply Shape-guided Cascade for Instance Segmentation},
  author    =  {Ding, Hao and Qiao, Siyuan and Yuille, Alan and Shen, Wei},
  booktitle =  {IEEE Conf. Comput. Vis. Pattern Recog. (CVPR)},
  year      =  {2021}
}

License

For academic use, this project is licensed under the 2-clause BSD License - see the LICENSE file for details. For commercial use, please contact Hao Ding (email) and Wei Shen.

hding2455/DSC