HandNet-Pipeline

A pipeline for hand detection and hand pose estimation using FCOS and A2J. Training and inference scripts are provided. Below is a demo of HandNet-Pipeline on ROS.

If you find this code useful for your research, consider citing the publication this pipeline was developed for:

@inproceedings{
    khargonkar2022neuralgrasps,
    title={NeuralGrasps: Learning Implicit Representations for Grasps of Multiple Robotic Hands},
    author={Ninad Khargonkar and Neil Song and Zesheng Xu and B Prabhakaran and Yu Xiang},
    booktitle={6th Annual Conference on Robot Learning},
    year={2022},
    url={https://openreview.net/forum?id=OzP68WT3UA}
}

Environment setup

Make sure to have Anaconda installed before initializing the environment.
Download MANO models and code (mano_v1_2.zip) from the MANO website, unzip, and place models/MANO_*.pkl files under misc/mano.
Run the following command to install the dependencies:
```
./scripts/init_env.sh
```
Activate the environment using the following command:
```
conda activate handnet_pipeline
```
Download the pretrained models using the following command:
```
./scripts/download_models.sh
```

ROS Demo

Make sure ROS environment variables are exported before running the demo and your active shell environment is bash.
Run the following command to start the demo (with pretrained models):
```
python ros_demo.py
```

Training

Detector Training

Download 100DOH (Pascal VOC format) and unzip into data directory
Run the following command to train the detector:
```
python trainval_net_fcos.py
```

Detector Testing

Run the following command to test the detector:

python trainval_net_fcos.py --test-only --resume=/path/to/checkpoint

A2J Training

Download DexYCB (subject 1 required), unzip, and run the following command

export DEX_YCB_DIR=/path/to/dex-ycb

$DEX_YCB_DIR should be a folder with the following structure:

├── 20200709-subject-01/
├── 20200813-subject-02/
├── ...
├── calibration/
└── models/

Run index preprocessing for DexYCB
```
python3 refine_idx_gen.py
```
Run the following command to train the A2J model:
```
python trainval_net_a2j.py fit
```

A2J Testing

Modify config/a2j.yaml to set the path to the trained A2J checkpoint.
Run the following command to test the A2J model:
```
python trainval_net_a2j.py test
```
If using pretrained model, load pretrained weights in LightningModule init instead (line 273).

Acknowledgements

Without these works, HandNet-Pipeline would not be possible. A huge thanks to the authors for their work and releasing their code.

FCOS

Code related to FCOS is adapted from the official Torchvision implementation

A2J

Code related to A2J is adapted from the official implementation

100 Days of Hands

Code related to 100 Days of Hands is adapted from the official implementation

DexYCB

Code related to DexYCB is adapted from the official implementation

Shaurya0108/handnet-pipeline