Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

This repo contains the supported code and configuration files to reproduce semantic segmentaion results of TransDA.

Paper

Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

Abstract After the great success of Vision Transformer variants (ViTs) in computer vision, it has also demonstrated great potential in domain adaptive semantic segmentation. Unfortunately, straightforwardly applying local ViTs in domain adaptive semantic segmentation does not bring in expected improvement. We find that the pitfall of local ViTs is due to the severe high-frequency components generated during both the pseudo-label construction and features alignment for target domains. These high-frequency components make the training of local ViTs very unsmooth and hurt their transferability. In this paper, we introduce a low-pass filtering mechanism, momentum network, to smooth the learning dynamics of target domain features and pseudo labels. Furthermore, we propose a dynamic of discrepancy measurement to align the distributions in the source and target domains via dynamic weights to evaluate the importance of the samples. After tackling the above issues, extensive experiments on sim2real benchmarks show that the proposed method outperforms the state-of-the-art methods.

Author

Runfa Chen, Yu Rong, Shangmin Guo, Jiaqi Han, Fuchun Sun, Tingyang Xu, Wenbing Huang

Citation

If you find this code useful for your research, please cite our paper:

@misc{chen2022smoothing,
    title={Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation},
    author={Runfa Chen and Yu Rong and Shangmin Guo and Jiaqi Han and Fuchun Sun and Tingyang Xu and Wenbing Huang},
    year={2022},
    eprint={2203.07988},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Prerequisites

Python 3.8
Pytorch 1.7.1
torchvision from master
yacs
tqdm
matplotlib
GCC >= 4.9
OpenCV
CUDA >= 10.1
mmcv >= 1.3.0
cityscapesscripts

Datasets

Generate the label statics file for GTA5 and SYNTHIA Datasets by running

python datasets/generate_gta5_label_info.py -d datasets/gta5 -o datasets/gta5/
python datasets/generate_synthia_label_info.py -d datasets/synthia -o datasets/synthia/

The data folder should be structured as follows:

├── datasets/
│   ├── cityscapes/     
|   |   ├── gtFine/
|   |   ├── leftImg8bit/
│   ├── gta5/
|   |   ├── images/
|   |   ├── labels/
|   |   ├── gtav_label_info.p
│   ├── synthia/
|   |   ├── RAND_CITYSCAPES/
|   |   ├── synthia_label_info.p
│   └── 			
...

Train & Validate & Test

We provide the script using 4 Tesla V100 GPUs.

TransDA-S

bash bin_small.sh

TransDA-B

bash cls_base.sh

Resume and test from our best checkpoint

checkpoint: model_last.pth (You need to save to './pth/CHECKPOINT_FOLDER_NAME/model_last.pth')
Swin ViT weight can be downloaded from Swin Transformer for ImageNet Classification and export weight = YOUR_WEIGHT_PATH.
TransDA-S checkpoint can be downloaded from GTA5toCityscapes and SYNTHIAtoCityscapes.

export base_dir=.
export dataset=_gta5
# export dataset=_synthia
export weight=https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_small_patch4_window7_224.pth
# validate
python test.py -cfg configs/train$dataset.yaml --saveres data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/val resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth MODEL.WEIGHTS $weight SOLVER.DIS binary
# test for submit
python test.py -cfg configs/train$dataset.yaml data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/test resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth  DATASETS.TEST cityscapes_test  MODEL.WEIGHTS $weight SOLVER.DIS binary

TransDA-B checkpoint can be downloaded from GTA5toCityscapes and SYNTHIAtoCityscapes.

export base_dir=.
export dataset=_gta5
# export dataset=_synthia
export weight=https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22k.pth
# validate
python test.py -cfg configs/train$dataset.yaml --saveres data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/val resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth MODEL.WEIGHTS $weight SOLVER.DIS class
# test for submit
python test.py -cfg configs/train$dataset.yaml data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/test resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth  DATASETS.TEST cityscapes_test  MODEL.WEIGHTS $weight SOLVER.DIS class

Visualization results

Acknowledgments

This project is based on FADA and Swin Transformer for Semantic Segmentaion. We thank them for their excellent open-source projects.

scottjingtt / TransDA