Rethinking Ensemble-Distillation for Semantic Segmentation Based Unsupervised Domain Adaptation

This repository includes the PyTorch implementation for the paper Rethinking Ensemble-Distillation for Semantic Segmentation Based Unsupervised Domain Adaptation.

[Video]

---

File Structure

weights/
├── weights/
|   ├── synthia/
|   ├── gta5/
|   |   ├── gta5_ours_drn_57.98.pth
|   |   ├── ...
Rethinking_EnD_UDA/
├── label_fusion/
├── train_deeplabv2/
├── train_deeplabv3+/
├── ...
Warehouse/
├── SYNTHIA/
│   ├── labels/
│   ├── images/
|   |   ├── 0000000.png
|   |   ├── 0000001.png
|   |   ├── ...
├── GTA5/
│   ├── image/
│   ├── labels/
|   |   ├── 00000.png
|   |   ├── 00001.png
|   |   ├── ...
├── Cityscapes/
│   ├── data/
│   │   ├── gtFine/
│   │   ├── leftImg8bit/
│   │   │   ├── train/
│   |   |   ├── val/
│   |   |   ├── test/
│   │   |   |   ├── aachen
│   │   |   |   ├── ...

Training

Quick Start:

1. Download the pre-generated pseudo labels here.
2. Place the pseudo labels in Cityscapes/data/gtFine folder and train the model with the following commands:

cd train_deeplabv3+
python train.py --class-balance --often-balance --backbone drn --restore-from ../../weights/weights/gta5/source/model_34.80.pth

The whole training procedure:

1. Train the teacher models

• DACS
• CRST
• CBST
• R-MRNet

2. Generate the pseudo labels and the output tensors. (NOTE: it is recommended that the certainty tensors should be first mapped to 0~100 and stored using byte tensors for memory conservation.)

3. Fuse the pseudo labels

cd label_fusion
python3 label_fusion.py

4. Place the pseudo labels in Cityscapes/data/gtFine folder and follow the instructions in "Quick Start" to train the model.

Testing

================ GTA5 ================
{ Deeplabv3+ }
cd train_deeplabv3+
python test.py --backbone drn --restore-from ../../weights/weights/gta5/gta5_ours_drn_57.98.pth

============== SYNTHIA ===============
{ Deeplabv3+ }
cd train_deeplabv3+
python test.py --num-classes 16 --source-domain synthia --backbone drn --restore-from ../../weights/weights/synthia/synthia_ours_drn_59.95.pth

Pretrained Weights

You can download the pre-trained weights here.

Prerequisites

• Python 3.6
• Pytorch 1.5.0

Download the dependencies:

pip install requirement.txt

Reference

If you find the code useful for your research, please consider citing

@InProceedings{Chao_2021_CVPR,
    author    = {Chao, Chen-Hao and Cheng, Bo-Wun and Lee, Chun-Yi},
    title     = {Rethinking Ensemble-Distillation for Semantic Segmentation Based Unsupervised Domain Adaption},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {2610-2620}
}

Acknowledgement

The code is partially borrowed from the following works:

• R-MRNet: https://github.com/layumi/Seg-Uncertainty
• Deeplabv3+: https://github.com/jfzhang95/pytorch-deeplab-xception