This is the official implementation of our ECCV 2022 paper "TACS: Taxonomy Adaptive Cross-Domain Semantic Segmentation"
Prerequisite
- CUDA/CUDNN
- Python3
- Packages found in requirements.txt
Dataset Preparation
GTAV, SYNTHIA, Cityscapes, Synscapes
Open Taxonomy Setting
Before Relabeling: python3 trainTACS_open_adduncertaincontrastive.py --config ./configs/configUDA_euler.json --name TACS_open_uncertaincontrast_beforerelabel --numsamples 30
After Relabeling: python3 trainTACS_open_adduncertaincontrastive_relabel.py --config ./configs/configUDA_euler_resume.json --name TACS_open_uncertaincontrast_afterrelabel --numsamples 30 --resume <Path to CheckPoint Before Relabeling>
Coarse-to-Fine Taxonomy Setting
Before Relabeling: python3 trainTACS_coarsetofine_adduncertaincontrastive.py --config ./configs/configUDA_euler.json --name TACS_coarsetofine_uncertaincontrast_beforerelabel --numsamples 30
After Relabeling: python3 trainTACS_coarsetofine_adduncertaincontrastive_relabel.py --config ./configs/configUDA_euler_resume.json --name TACS_coarsetofine_uncertaincontrast_afterrelabel --numsamples 30 --resume <Path to CheckPoint Before Relabeling>
Implicitly Overlapping Taxonomy Setting
Before Relabeling: python3 trainTACS_implicitoverlapping_adduncertaincontrastive.py --config ./configs/configUDA_euler.json --name TACS_implicitoverlapping_uncertaincontrast_beforerelabel --numsamples 15
After Relabeling: python3 trainTACS_implicitoverlapping_adduncertaincontrastive_relabel.py --config ./configs/configUDA_euler_resume.json --name TACS_implicitoverlapping_uncertaincontrast_afterrelabel --numsamples 15 --resume <Path to CheckPoint Before Relabeling>
Model Testing
If you would like to drectly use our trained model, please access our checkpoints folder.
Open, Coarse-to-Fine: python3 evaluateTACS.py --model-path <Path to Checkpoint>
Implicitly-Overlapping: python3 evaluateTACS_16classes.py --model-path <Path to Checkpoint>
Acknowledgements
The implementation is based on the following open-source projects. We thank their authors for making the source code publicly available.
Citation
If this helps you, please cite our TACS work:
@inproceedings{gong2022tacs,
title={TACS: Taxonomy Adaptive Cross-Domain Semantic Segmentation},
author={Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, Luc Van Gool},
booktitle={ECCV},
year={2022}
}