Home

Awesome

Balancing Logit Variation for Long-tailed Semantic Segmentation, CVPR 2023

Fully-supervised semantic segmentation

Installation

cd fully_sup
conda create -n blv python=3.7 -y
conda activate blv
conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge
pip install openmim
mim install mmcv-full==1.4.0
pip install -e .

Data Preparation

Please follow this link dataset_prepare.md to setup the datasets.

Run

For some models, you should download the corresponding pretrained checkpoints for the backbone manually.

cd fully_sup
python -u tools/train.py /path/to/the/config/file

Results and configs

MethodBackbonemIoUmIoU(tail)config
HRNet-18OCRHead79.2263.51config
HRNet-18+BLV79.9466.70config
ResNet50UperHead78.2862.56config
ResNet50+BLV78.6364.57config
ResNet50PSPHead77.9861.96config
ResNet50+BLV78.5363.34config
ResNet101UperHead79.4164.68config
ResNet101+BLV79.8866.29config
MiT-b0SegformerHead76.8567.58config
MiT-b0+BLV77.0968.91config
Swin-TK-NeT79.6871.70config
Swin-T+BLV80.1172.94config
Vit-B16K-NeT76.4868.25config
Vit-B16+BLV77.6870.63config

Unsupervised Domain adaptive semantic segmentation

Semi-supervised semantic segmentation

Citation

If you find this useful for your research, please cite the following paper.

@inproceedings{wang2023balancing,
  title={Balancing Logit Variation for Long-tailed Semantic Segmentation},
  author={Wang, Yuchao and Fei, Jingjing and Wang, Haochen and Li, Wei and Bao, Tianpeng and Wu, Liwei and Zhao, Rui and Shen, Yujun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19561--19573},
  year={2023}
}

Acknowledgements

The implementation of fully-supervised semantic segmentation task is based on mmsegmentation.

The implementation of unsupervised domain adaptived semantic segmentation task is based on HRDA.