This is the official implementation for:
Faster VoxelPose: Real-time 3D Human Pose Estimation by Orthographic Projection,
Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang
ECCV 2022
conda install pytorch torchvision cudatoolkit=<your cuda version>
pip install -r requirements.txt
We use the Shelf/Campus and CMU Panoptic datasets. Please refer to VoxelPose for detailed instructions.
The directory tree should look like this:
${ROOT}
|-- data
|-- Panoptic
|-- 16060224_haggling1
| |-- hdImgs
| |-- hdvideos
| |-- hdPose3d_stage1_coco19
| |-- calibration_160224_haggling1.json
|-- 160226_haggling1
|-- ...
|-- Shelf
| |-- Camera0
| |-- ...
| |-- Camera4
| |-- actorsGT.mat
| |-- calibration_shelf.json
| |-- pred_shelf_maskrcnn_hrnet_coco.pkl
|-- Campus
| |-- Camera0
| |-- Camera1
| |-- Camera2
| |-- actorsGT.mat
| |-- calibration_campus.json
| |-- pred_campus_maskrcnn_hrnet_coco.pkl
|-- panoptic_training_pose.pkl
Train and validate on the five selected camera views. You can specify the GPU devices and batch size per GPU in the config file.
python run/train.py --cfg configs/panoptic/jln64.yaml
python run/train.py --cfg configs/shelf/jln64.yaml
python run/train.py --cfg configs/campus/jln64.yaml
Evaluate the models. It will print evaluation results to the screen./
python run/validate.py --cfg configs/panoptic/jln64.yaml
It will print the PCP results to the screen.
python run/validate.py --cfg configs/shelf/jln64.yaml
python run/validate.py --cfg configs/campus/jln64.yaml
If you use our code or models in your research, please cite with:
@inproceedings{fastervoxelpose,
author={Ye, Hang and Zhu, Wentao and Wang, Chunyu and Wu, Rujie and Wang, Yizhou},
title={Faster VoxelPose: Real-time 3D Human Pose Estimation by Orthographic Projection},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2022}
}