Awesome

Point2Mask: Point-supervised Panoptic Segmentation via Optimal Transport

Wentong Li, Yuqian Yuan, Song Wang, Jianke Zhu, Jianshu Li, Jian Liu, and Lei Zhang

Paper (arXiv). ICCV2023.

Environment Setup

conda create -n point2mask python=3.8 -y
conda activate point2mask
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install openmim
mim install mmdet==2.17.0
mim install mmcv-full==1.3.9
git clone https://github.com/LiWentomng/Point2Mask.git
cd Point2Mask
pip install -r requirements.txt
sh setup.sh

Model Zoo

1.Single-point Supervision(P1)

Pascal VOC

Backbone	Supervision	Models	PQ	PQ_th	PQ_st
ResNet-50	P1	model	53.7	51.9	90.5
ResNet-101	P1	model	54.8	53.0	90.4
Swin-L	P1	model	61.0	59.4	93.0

COCO

Backbone	Supervision	Models	PQ	PQ_th	PQ_st
ResNet-50	P1	model	32.4	32.6	32.2
ResNet-101	P1	model	34.0	34.3	33.5
Swin-L	P1	model	37.0	37.0	36.9

Pascal VOC with COCO P1-pretrained model.

Backbone	Supervision	Models	PQ	PQ_th	PQ_st
ResNet-50	P1	model	60.7	59.1	91.8
ResNet-101	P1	model	63.2	61.8	92.3
Swin-L	P1	model	64.2	62.7	93.2

2.Ten-point Supervision(P10)

Pascal VOC

Backbone	Supervision	Models	PQ	PQ_th	PQ_st
ResNet-50	P10	model	59.1	57.5	91.8
ResNet-101	P10	model	60.2	58.6	92.1

COCO

Backbone	Supervision	Models	PQ	PQ_th	PQ_st
ResNet-50	P10	model	32.4	32.6	32.2
ResNet-101	P10	model	36.7	37.3	35.7

Get Started

We use Pascal VOC and COCO datasets, please see Preparing datasets for Point2Mask.

Demo

To test our model with an input image, please run demo.py

python  demo.py --config-file ./configs/point2mask/voc/point2mask_voc_wsup_r50.py --weights /path/to/coco_r50.pth --input image.jpg --out-file prediction.jpg

Training

For VOC training:

CUDA_VISIBLE_DEVICE=0,1,2,3  tools/dist_train.sh configs/point2mask/voc/point2mask_voc_wsup_r50.py 4

For COCO training:

CUDA_VISIBLE_DEVICE=0,1,2,3,4,5,6,7  tools/dist_train.sh configs/point2mask/coco/point2mask_coco_wsup_r50.py 8

Note: our models for Pascal VOC are trained with 3090/V100 gpus, and for COCO with A100 GPUs. The Structured Edge (SE) model used for low-level edege is here.

Test

For PQ evaluation:

CUDA_VISIBLE_DEVICE=0,1,2,3  tools/dist_test.sh configs/point2mask/voc/point2mask_voc_wsup_r50.py work_dirs/xxx.pth 4  --eval pq

For visual results:

CUDA_VISIBLE_DEVICE=0 tools/dist_test.sh configs/point2mask/voc/point2mask_voc_wsup_r50.py work_dirs/xxx.pth 1 --show-dir xxx

Visual Examples

Visual Results on COCO with ResNet-101. <img src="./data/visualization.jpg" width="800px">

Visualization of the learning high-level boundary map. <img src="./data/highlevel_bundary_map.png" width="800px">

Acknowledgement

Code is largely based on PSPS, Panoptic Segformer, MMdetection.
Thanks for their great open-source projects!

Citation

@inproceedings{point2mask,
  title={Point2Mask: Point-supervised Panoptic Segmentation via Optimal Transport},
  author={Wentong Li, Yuqian Yuan, Song Wang, Jianke Zhu, Jianshu Li, Jian Liu and Lei Zhang},
  booktitle={Proceedinngs of IEEE International Conference on Computer Vision},
  year={2023}
}