Official Implementation for: SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation
[SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation],
Sichen Chen*, Yingyi Zhang*, Siming Huang*, Ran Yi, Ke Fan, Ruixin Zhang, Peixia Chen, Jun Wang, Shouhong Ding, Lizhuang Ma.
In: Conference on Computer Vision and Pattern Recognition~(CVPR), 2024
arXiv preprint (arXiv 2404.03518)
(* equal contribution)
Most of the top-performing transformer-based human pose estimation models are too computation-consuming and storage-demanding to deploy on edge computing platforms. Those transformer-based models that require fewer resources are prone to under-fitting due to their smaller scale and thus perform notably worse than their larger counterparts. Given this conundrum, we introduce SDPose, a new self-distillation method for improving the performance of small transformer-based models.
SDPose referenced DistilPose (CVPR 2023) and is developed using MMPose and Pytorch framework. Please install the relevant packages listed below:
conda create -n sdpose python=3.8 pytorch=1.7.0 torchvision -c pytorch -y
conda activate sdpose
pip3 install openmim
mim install mmcv-full==1.3.8
git submodule update --init
cd mmpose
git checkout v0.29.0
pip3 install -e .
cd ..
pip3 install -r requirements.txt
For training on COCO, you will need to download the official COCO dataset and modify the dataset path in the model configuration files. After these setups, run the following command lines:
./tools/dist_train.sh configs/body/2d_kpt_sview_rgb_img/sdpose/coco/sdpose_s_v1_stemnet_coco_256x192.py 8
For evaluating on COCO, downlowd checkpoint and run the following command lines:
./tools/dist_test.sh configs/body/2d_kpt_sview_rgb_img/distilpose/coco/sdpose_s_v1_stemnet_coco_256x192.py \
./checkpoints/sdpose_s_v1.pth 8
| Name | Param | GFLOPs | AP | download |
|---|---|---|---|---|
| SDPose-T | 4.4 | 1.8 | 69.7 | Download |
| SDPose-S-V1 | 6.6 | 2.4 | 72.3 | Download |
| SDPose-S-V2 | 6.2 | 4.7 | 73.5 | Download |
| SDPose-B | 13.2 | 5.2 | 73.7 | Download |
Please consider citing our papers in your publications if the project helps your research. BibTeX reference is as follows.
@article{chen2024sdpose,
title={SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation},
author={Chen, Sichen and Zhang, Yingyi and Huang, Siming and Yi, Ran and Fan, Ke and Zhang, Ruixin and Chen, Peixian and Wang, Jun and Ding, Shouhong and Ma, Lizhuang},
journal={arXiv preprint arXiv:2404.03518},
year={2024}
}