Awesome

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

Fully-supervised semantic segmentation.
Unsupervised Domain adaptive semantic segmentation.
Semi-supervised semantic segmentation.

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

Method	Backbone	mIoU	mIoU(tail)	config
HRNet-18	OCRHead	79.22	63.51	config
HRNet-18	+BLV	79.94	66.70	config
ResNet50	UperHead	78.28	62.56	config
ResNet50	+BLV	78.63	64.57	config
ResNet50	PSPHead	77.98	61.96	config
ResNet50	+BLV	78.53	63.34	config
ResNet101	UperHead	79.41	64.68	config
ResNet101	+BLV	79.88	66.29	config
MiT-b0	SegformerHead	76.85	67.58	config
MiT-b0	+BLV	77.09	68.91	config
Swin-T	K-NeT	79.68	71.70	config
Swin-T	+BLV	80.11	72.94	config
Vit-B16	K-NeT	76.48	68.25	config
Vit-B16	+BLV	77.68	70.63	config

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.