FPR: False Positive Rectification for Weakly Supervised Semantic Segmentation [pdf]

Method

The CAM in weakly supervised semantic segmentation suffers from over-activation from co-occurred background regions. Luckily, We observe that co-occurred background sometimes present in image solely and its cues can be easily captured through image-level labels. In detail, the framework is as follows, which insists of Online Prototype Computing and Training with Prototype parts.

Get Start

Environment

git clone https://github.com/mt-cly/FPR.git
conda create -n fpr python=3.8
conda activate fpr
cd ./FPR
pip install -r requirements.txt

Datasets & Weights

1. Download the augmented VOC12 & COCO14 and put them to the Dataset folder.
2. Download the CAM initial weight and place it to the sess folder.

Run

python run_sample.py

The trained fpr weight and log are available: [res50_fpr.pth] [fpr.log]

TODO

• Release the COCO2014 part code.

Citation

@article{chen2023fpr,
  title={FPR: False Positive Rectification for Weakly Supervised Semantic Segmentation},
  author={Chen, Liyi and Lei, Chenyang and Li, Ruihuang and Li, Shuai and Zhang, Zhaoxiang and Zhang, Lei},
  booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
  pages={},
  year={2023}
}