cheny124800 / CIRKD

[CVPR-2022] Official implementations of CIRKD: Cross-Image Relational Knowledge Distillation for Semantic Segmentation

Cross-Image Relational Knowledge Distillation for Semantic Segmentation

This repository contains the source code of CIRKD (Cross-Image Relational Knowledge Distillation for Semantic Segmentation).

Requirement

Ubuntu 18.04 LTS

Python 3.8 (Anaconda is recommended)

CUDA 11.1

PyTorch 1.8.0

NCCL for CUDA 11.1

Backbones pretrained on ImageNet:

• resnet101-imagenet.pth
• resnet18-imagenet.pth
• mobilenetv2-imagenet.pth

Performance on Cityscapes

All models are trained over 8 * NVIDIA GeForce RTX 3090

Role	Network	Method	Val mIoU	test mIoU	Pretrained	train script
Teacher	DeepLabV3-ResNet101	-	78.07	77.46	Google Drive	sh
Student	DeepLabV3-ResNet18	Baseline	74.21	73.45	-	sh
Student	DeepLabV3-ResNet18	CIRKD	76.38	75.05	Google Drive	sh
Student	DeepLabV3-ResNet18*	Baseline	65.17	65.47	-	sh
Student	DeepLabV3-ResNet18*	CIRKD	68.18	68.22	Google Drive	sh
Student	DeepLabV3-MobileNetV2	Baseline	73.12	72.36	-	sh
Student	DeepLabV3-MobileNetV2	CIRKD	75.42	74.03	Google Drive	sh
Student	PSPNet-ResNet18	Baseline	72.55	72.29	-	sh
Student	PSPNet-ResNet18	CIRKD	74.73	74.05	Google Drive	sh

*denotes that we do not initialize the backbone with ImageNet pre-trained weights.

Evaluate pre-trained models on Cityscapes val and test sets

Evaluate the pre-trained models on val set

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
  python -m torch.distributed.launch --nproc_per_node=8 eval.py \
  --model deeplabv3 \
  --backbone resnet101 \
  --data [your dataset path]/cityscapes/ \
  --save-dir [your directory path to store log files] \
  --pretrained [your checkpoint path]/deeplabv3_resnet101_citys_best_model.pth

Generate the resulting images on test set

CUDA_VISIBLE_DEVICES=0,1,2,3 \
  python -m torch.distributed.launch --nproc_per_node=4 test.py \
  --model deeplabv3 \
  --backbone resnet101 \
  --data [your dataset path]/cityscapes/ \
  --save-dir [your directory path to store resulting images] \
  --save-pred \
  --pretrained [your checkpoint path]/deeplabv3_resnet101_citys_best_model.pth

You can submit the resulting images to the Cityscapes test server.

The current codes have been reorganized and we have not tested them thoroughly. If you have any questions, please contact us without hesitation.

Citation

@inproceedings{yang2022cross,
  title={Cross-Image Relational Knowledge Distillation for Semantic Segmentation},
  author={Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, Qian Zhang},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}