A Dual-Branch Self-Boosting Framework for Self-Supervised 3D Hand Pose Estimation (TIP2022)

This is the official PyTorch implementation code. For technical details, please refer to:

A Dual-Branch Self-Boosting Framework for Self-Supervised 3D Hand Pose Estimation
Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, Jianxin Liao
[Paper]

3D Hand Pose Estimation and 3D Hand Mesh Reconstruction

Compared with semi-automatic annotation methods in ICVL and MSRA datasets, our self-supervised method can generate more accurate and robust 3D hand pose and hand mesh.

ICVL Dataset

demo1

MSRA Dataset

demo2

Skeleton-based Action Recognition

Using the 3D skeleton generated by DSF can greatly improve the accuracy of the skeleton-based action recognition.

Method	Modality	SHREC 14	SHREC 28	DHG 14	DHG 28
PointLSTM	Point clouds	95.9	94.7	-	-
Res-TCN	Skeleton	91.1	87.3	86.9	83.6
ST-GCN	Skeleton	92.7	87.7	91.2	87.1
STA-Res-TCN	Skeleton	93.6	90.7	89.2	85.0
ST-TS-HGR-NET	Skeleton	94.3	89.4	87.3	83.4
HPEV	Skeleton	94.9	92.3	92.5	88.9
DG-STA	Skeleton	94.4	90.7	91.9	88.0
DG-STA (AWR)	Skeleton	96.3_↑1.9	93.3_↑2.6	94.5_↑2.6	92.1 _↑4.1
DG-STA (DSF)	Skeleton	96.8_↑2.4	95.0_↑4.3	96.3_↑4.4	95.9_↑7.9

Installation

Prerequisites

Python >= 3.8
PyTorch >= 1.10
pytorch3d == 0.4.0
CUDA (tested with cuda11.3)
Other dependencies described in requirements.txt

MANO

Go to MANO website
Download Models and copy the models/MANO_RIGHT.pkl into the MANO folder
Your folder structure should look like this:
```
DSF/
MANO/
MANO_RIGHT.pkl
```
NYU Dataset
Download and decompress NYU and modify the root_dir in config.py according to your setting.
Download the center files [Google Drive] and put them into the train and test directories of NYU respectively.
Download the MANO parameter file of the NYU dataset we generated MANO Files

Your folder structure should look like this:

.../
nyu/
train/
  center_train_0_refined.txt
  center_train_1_refined.txt
  center_train_2_refined.txt
  ...
test/
  center_test_0_refined.txt
  center_test_1_refined.txt
  center_test_2_refined.txt
  ...
posePara_lm_collosion/
  nyu-train-0-pose.txt
  ...

Pretrained Model

Download our pre-trained model with self-supervised training [Google Drive]
Download our pre-trained model with only synthetic data [Google Drive]
Download the Consis-CycleGAN model [Google Drive]

Running DSF

Evaluation

Set load_model as the path to the pretrained model and change the phase to "test" in config.py, run

python train_render.py

Self-supervised Training

To perform self-supervised training, set finetune_dir as the path to the pretrained model with only synthetic data and tansferNet_pth as the path to the Consis-CycleGAN model in config.py. Then, change the phase to "train", run

python train_render.py

Pre-training with Synthetic Data

To perform pre-training, set train_stage to "pretrain" in config.py, run

python train_render.py

Citation

If you find our work useful in your research, please citing:

@ARTICLE{9841448,
  author={Ren, Pengfei and Sun, Haifeng and Hao, Jiachang and Qi, Qi and Wang, Jingyu and Liao, Jianxin},
  journal={IEEE Transactions on Image Processing}, 
  title={A Dual-Branch Self-Boosting Framework for Self-Supervised 3D Hand Pose Estimation}, 
  year={2022},
  volume={31},
  number={},
  pages={5052-5066},
  doi={10.1109/TIP.2022.3192708}}

PengfeiRen96 / DSF

readme