T2P

Official repository of: Multi-agent Long-term 3D Human Pose Forecasting via Interaction-aware Trajectory Conditioning (CVPR 2024 Highlight) [paper] [project page]

Table of Contents

• Updates
• Dataset preparation (JRDB)
• Acknowledgements

Dataset

We upload an updated version of the JRDB dataset parser (3D joints => SMPL parameters for pose).

1. Set conda environment based on BEV. Also install torch-geometric. (Tested version: Python 3.11.9, cuda 12.1, torch 2.3.0, torch-geometric 2.5.3)
2. Download the original dataset from JRDB. Change the default_save_dir in preprocess_1st_jrdb.py accordingly.
3. Download the preprocessed robot odometry files (.npy) from releases. Change the directory of odometry_base in preprocess_1st_jrdb.py accordingly. Thanks to human-scene-transformer for sharing your work.
4. Process the 1st and 2nd in order (preprocess_1st_jrdb.py, preprocess_2nd_jrdb.py).

preprocess_1st_jrdb.py: Processes the trajectory information from 3D bounding boxes, and 3D pose is extracted by BEV. Theta parameters of SMPL (24X3) is used as pose information. Each frame is preprocessed independently. preprocess_2nd_jrdb.py: Parses each scene into .pt file. The data is saved as TemporalData class, a format used by HiVT. The parameters are set to parse the data in 2.5FPS.

Updates

[2024-10-17]

-JRDB Dataset parser uploaded: Codes to model to come shortly!

[2024-06-16]

-Project page created: Project page up and running. Codes & dataset to come shortly!

Acknowledgements

• Thanks to human-scene-transformer, JRDB, BEV, HiVT for sharing your work.