The official implementation of paper:
If you find this repository useful, please cite:
@InProceedings{Wang_2021_ICCV,
author = {Wang, Jian and Liu, Lingjie and Xu, Weipeng and Sarkar, Kripasindhu and Theobalt, Christian},
title = {Estimating Egocentric 3D Human Pose in Global Space},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2021},
pages = {11500-11509}
}
- Install pytorch 1.4+ with cuda support.
- Run
mkdir networks/logs
and download the trained VAE model into directorynetworks/logs
from here. - Run
mkdir data
and download the processed test sequences into directorydata
from here. - Run the test on the sequences:
python optimize_whole_sequence.py --data_path data/jian3
python optimize_whole_sequence.py --data_path data/studio-jian1
python optimize_whole_sequence.py --data_path data/studio-jian2
python optimize_whole_sequence.py --data_path data/studio-lingjie1
python optimize_whole_sequence.py --data_path data/studio-lingjie2
If you want to train the motion VAE, See directory networks
If you want to run on your own dataset,
you need to firstly preprocess the data following the scripts in directory: MakeDataForOptimization
.
Basically you need to prepare the following data:
- Predicted human body heatmap in egocentric view.
- Predicted human body joint depths in egocentric view.
- Human body ground truth (for calculating MPJPEs).
- Camera pose sequence from the OpenVSLAM method.
All of these data are combined into a single pickle file for each sequence.