S2C

Official repository for CVPR 2024 Oral paper: "From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation" by Hyeokjun Kweon.

Prerequisite

Tested on Ubuntu 18.04, with Python 3.8, PyTorch 1.8.2, CUDA 11.4, 4 GPUs.
The PASCAL VOC 2012 development kit: You need to specify place VOC2012 under ./data folder.
ImageNet-pretrained weights for resnet38d are from [resnet_38d.params]. You need to place the weights as ./pretrained/resnet_38d.params.

Prerequisite on SAM

Please install SAM and download vit_h version as ./pretrained/sam_vit_h.pth
Note that I slightly modified the original code of SAM for fast batch-wise inference during the training of CAMs.
After installing SAM properly, you should substitute the files 'mask_decoder.py' and 'sam.py' in the segment_anything/modeling directory with the files in 'modeling' of this repository.
Additionally, you need to run the Segment-Everything option using SAM as preprocessing. Please refer to get_se_map.py for further details.

Usage

This repository generates CAMs (seeds) to train the segmentation network.
For further refinement, refer RIB and SAM_WSSS.

Training

Please specify the name of your experiment.
Training results are saved at ./experiment/[exp_name]
```
python train.py --name [exp_name] --model s2c
```

Evaluation for CAM

python evaluation.py --name [exp_name] --task cam --dict_dir dict

Citation

If our code be useful for you, please consider citing our CVPR 2024 paper using the following BibTeX entry.

@inproceedings{kweon2024sam,
  title={From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation},
  author={Kweon, Hyeokjun and Yoon, Kuk-Jin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19499--19509},
  year={2024}
}

sangrockEG / S2C