This repo provides details about how to use SOLIDER pretrained representation on human parsing task. We modify the code from Self-Correction-Human-Parsing, and you can refer to the original repo for more details.

Details of installation and dataset preparation can be found in Self-Correction-Human-Parsing.

Step 1. Download models from SOLIDER, or use SOLIDER to train your own models.

Steo 2. Put the pretrained models under the pretrained file, and rename their names as ./pretrained/solider_swin_tiny(small/base).pth

Train with single GPU or multiple GPUs:

sh train_swin.sh

• We use the pretrained models from SOLIDER.
• The semantic weight we used in these experiments is 0.8.

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

@inproceedings{chen2023beyond,
  title={Beyond Appearance: a Semantic Controllable Self-Supervised Learning Framework for Human-Centric Visual Tasks},
  author={Weihua Chen and Xianzhe Xu and Jian Jia and Hao Luo and Yaohua Wang and Fan Wang and Rong Jin and Xiuyu Sun},
  booktitle={The IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023},
}