Home

Awesome

Interacting Two-Hand 3D Pose and Shape Reconstruction from Single Color Image

Interacting Two-Hand 3D Pose and Shape Reconstruction from Single Color Image

Baowen Zhang, Yangang Wang, Xiaoming Deng*, Yinda Zhang*, Ping Tan, Cuixia Ma and Hongan Wang

Project page       Paper       Supp

prediction example

This repository contains the model of the ICCV'2021 paper "Interacting Two-Hand 3D Pose and Shape Reconstruction from Single Color Image".

We propose a novel deep learning framework to reconstruct 3D hand poses and shapes of two interacting hands from a single color image. Previous methods designed for single hand cannot be easily applied for the two hand scenario because of the heavy inter-hand occlusion and larger solution space. In order to address the occlusion and similar appearance between hands that may confuse the network, we design a hand pose-aware attention module to extract features associated to each individual hand respectively. We then leverage the two hand context presented in interaction and propose a context-aware cascaded refinement that improves the hand pose and shape accuracy of each hand conditioned on the context between interacting hands. Extensive experiments on the main benchmark datasets demonstrate that our method predicts accurate 3D hand pose and shape from single color image, and achieves the state-of-the-art performance.

Update (2023-2-10)

Training and testing codes are released!

The new version with training and testing codes is at Two-Hand-Shape-Pose, which was trained on InterHand2.6M(v1.0).

1. Installation

This code is tested with Cuda 11.1.

Clone this repository.

git clone https://github.com/BaowenZ/Two-Hand-Shape-Pose.git
cd Two-Hand-Shape-Pose

In the following, ${TWO_HAND} refers to Two-Hand-Shape-Pose.

Install dependencies

conda create -n intershape python=3.9
conda activate intershape
pip install --upgrade pip
pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

2. Download models

Download pre-trained model model.pts and put it into folder model/.

Download the MANO model files from MANO. Unzip mano_v1_2.zip under ${TWO_HAND} and rename the unzipped folder as mano/.

3. Running the code

python test.py --test_folder test_data --model_path model/model.pts

Our model predicts hand meshes from images in test_data/. The estimated meshes are saved as obj files in test_data/.

4. A Note on Evaluation

Our model is trained and tested on InterHand2.6M v0 dataset. We use color images with MANO annotations to train our model. Samples without middle finger's MCP joint or root joint are not used during training and testing because we use these joints for alignment. Model trained on InterHand2.6M v1 will be released in the future.

5. Citation

Please consider citing the paper if you use this code.

@inproceedings{Zhang2021twohand, 
      title={Interacting Two-Hand 3D Pose and Shape Reconstruction from Single Color Image}, 
      author={Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma and Hongan Wang}, 
      booktitle={International Conference on Computer Vision (ICCV)}, 
      year={2021} 
} 

6. Acknowledgement

We use part of the great code from InterNet and mano layer.

Image samples in test_data/ are from InterHand2.6M.

We thank the authors of InterNet, InterHand2.6M and mano layer for their great work.

7. Contact Information

For any questions, feel free to contact: 947976219@qq.com, xiaoming@iscas.ac.cn