This project was undertaken as a part of COMS 6901 Projects in Computer Science at Columbia University, under Prof. Peter K Allen.
Implementation of H+O: Unified Egocentric Recognition of 3D Hand-Object Poses and Interactions
This repository provides the code for training a model that can jointly predict hand pose, object pose, object class, and the performed action.
Disclaimer: I have actually skipped the RNN part in the implementation. With a little effort it can be done.
- PyTorch
- tqdm
- tensorboardX
- torchvision
- trimesh
- matplotlib
This has been tested on Python 2.7, Ubuntu 16.04, and Pytorch 1.0.1.post2.
UnifiedPoseEstimation
- cfg
- data
- models
- unified_pose_estimation
If any of the above directory is not present, please create them.
You need to download and place the FPHA dataset in data directory. The data directory structure will look like
- Hand_pose_annotation_v1
- Object_models
- Subjects_info
- Video_files
- ..etc
Now cd into unified_pose_estimation
- python clean.py
- python train.py
python test.py