SOLIDER on [Human Pose]

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

Installation and Datasets

Details of installation and dataset preparation can be found in mmpose-installation.

Prepare Pre-trained Models

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

Training

Train with single GPU or multiple GPUs:

sh run_train.sh

Performance

Method	Model	COCO(AP/AR)
SOLIDER	Swin Tiny	74.4/79.6
SOLIDER	Swin Small	76.3/81.3
SOLIDER	Swin Base	76.6/81.5

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

Citation

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},
}