RMRKD: Region-aware mutual relational knowledge distillation for semantic segmentation

Environment

1. Clone our repo and create conda environment.

git clone https://github.com/Debrove/RMRKD.git && cd RMRKD
conda create -n rmrkd python=3.8
conda activate rmrkd

2. Install Pytorch and other dependencies Please refer MMSegmentation for detail installation.

pip install torch==1.12.0+cu116 torchvision==0.13.0+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
pip install openmim
mim install mmcv-full==1.7.0
mim install mmsegmentation==0.29.1
pip install -r requirements.txt

Dataset

Please follow MMSegmentation to prepare datasets.

Train

#single GPU
python tools/train.py configs/distillers/rmrkd/psp_r101_distill_psp_r18_40k_512x512_city.py

#multi GPU
bash tools/dist_train.sh configs/distillers/rmrkd/psp_r101_distill_psp_r18_40k_512x512_city.py 8

Transfer

# Tansfer the RMRKD model into mmseg model
python pth_transfer.py --mgd_path $ckpt --output_path $new_mmseg_ckpt

Test

#single GPU
python tools/test.py configs/pspnet/pspnet_r18-d8_512x512_40k_cityscapes.py $new_mmseg_ckpt --eval mIoU

#multi GPU
bash tools/dist_test.sh configs/pspnet/pspnet_r18-d8_512x512_40k_cityscapes.py $new_seg_ckpt 8 --eval mIoU

Results on CityScapes

Teacher	Student	Baseline(mIoU)	+RMRKD(mIoU)	config
PspNet-R101	PspNet-R18	69.37	75.72	config
PspNet-R101	DeepLabV3-R18	73.37	76.72	config
DeepLabV3 plus-R101	MobileNetV2	73.76	76.87	config
DeepLabV3-R101	MobileNetV2	73.11	76.32	config

Results on Pascal VOC

Teacher	Student	Baseline(mIoU)	+RMRKD(mIoU)	config
PspNet-R101	PspNet-R18	70.52	74.64	config
PspNet-R101	DeepLabV3-R18	71.60	74.97	config

Acknowledgements

Our code is based on MMSegmentation, MGD, CIRKD. Many thanks to these great works and open-source codebases.

Citation

@article{zheng2025region,
  title={Region-aware mutual relational knowledge distillation for semantic segmentation},
  author={Zheng, Haowen and Lin, Xuxin and Liang, Hailun and Zhou, Benjia and Liang, Yanyan},
  journal={Pattern Recognition},
  volume={161},
  pages={111319},
  year={2025},
  publisher={Elsevier}