Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation

Our paper is accepted by CVPR-2024

Picture: Overview of the proposed Single-to-Dual-view framework for adapting a traditional single-view hand pose estimator to arbitrary dual views.

Picture: The proposed architecture.

---

Results

This repository contains the official PyTorch implementation of the following paper:

Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation
Ruicong Liu, Takehiko Ohkawa, Mingfang Zhang, and Yoichi Sato

Abstract: The pursuit of accurate 3D hand pose estimation stands as a keystone for understanding human activity in the realm of egocentric vision. The majority of existing estimation methods still rely on single-view images as input, leading to potential limitations, e.g., limited field-of-view and ambiguity in depth. To address these problems, adding another camera to better capture the shape of hands is a practical direction. However, existing multi-view hand pose estimation methods suffer from two main drawbacks: 1) Requiring multi-view annotations for training, which are expensive. 2) During testing, the model becomes inapplicable if camera parameters/layout are not the same as those used in training. In this paper, we propose a novel Single-to-Dual-view adaptation (S2DHand) solution that adapts a pre-trained single-view estimator to dual views. Compared with existing multi-view training methods, 1) our adaptation process is unsupervised, eliminating the need for multi-view annotation. 2) Moreover, our method can handle arbitrary dual-view pairs with unknown camera parameters, making the model applicable to diverse camera settings. Specifically, S2DHand is built on certain stereo constraints, including pair-wise cross-view consensus and invariance of transformation between both views. These two stereo constraints are used in a complementary manner to generate pseudo-labels, allowing reliable adaptation. Evaluation results reveal that S2DHand achieves significant improvements on arbitrary camera pairs under both in-dataset and cross-dataset settings, and outperforms existing adaptation methods with leading performance.

Resources

Material related to our paper is available via the following links:

• Paper: https://arxiv.org/pdf/2403.04381.pdf
• Code: https://github.com/ut-vision/S2DHand
• AssemblyHands dataset: https://assemblyhands.github.io/
• Pre-trained models: https://drive.google.com/file/d/1rn-ez1rBjELTgcx1hU4gXgkH41ezBJbK/view?usp=sharing

System requirements

• Only Linux is tested.
• 64-bit Python 3.10 installation.

Playing with pre-trained networks and training

Data preparation

Please download the pre-trained models and AssemblyHands dataset first. Assuming the pre-trained models and dataset are stored under ${DATA_DIR}. The structure of ${DATA_DIR} follows

- ${DATA_DIR}
    - AsseblyHands
        - annotations
        - images      
    - S2DHand-pretrain 
        - ckp_detnet_37.pth
        - ckp_detnet_68.pth

Please run the following command to register the pre-trained model and data.

cp ${DATA_DIR}/S2DHand-pretrain/* pretrain
mkdir data
mkdir data/assemblyhands
ln -s ${DATA_DIR}/AssemblyHands/* data/assemblyhands

Optional: This code will automatically check all invalid data when first run, which takes lots of time. To save time, you can download the pre-generated invalid.txt here and move it to ${DATA_DIR}/AssemblyHands/annotations/${split}. For example, move invalid_train_ego.txt to ${DATA_DIR}/AssemblyHands/annotations/train.

Single-to-Dual-View Adaptation (Training)

run.sh provides a complete procedure of training and testing. The adapt_detnet_dual.py automatically tests the adapted model after each epoch of adaptation.

We provide several arguments for adaptation:

-trs and -tes. They represent training and testing dataset, please set it to "ah" (assemblyhands).

--root_idx represents the root joint index for alignment. Remember to set it as 0 for in-dataset adaptation while 9 for cross-dataset. This is because the pre-trained models are trained under different root_idxs.

--pic number of image pairs for adaptation or evaluation, set to --pic=-1 for using all image pairs..

--checkpoint directory to save models or load the model for evaluation.

--resume remember to add this argument for adaptation from an initial model.

--evaluate please use this argument of testing/evaluating an existing model.

-eid, --evaluate_id determines which epoch's model is used for adaptation or evaluation. Note that the code will first look for the model in --checkpoint.

--epochs number of adaptation epochs.

--start_epoch start epoch for saving models.

--setup can be set as 0 or 1, which stands for the id of the two headsets from AssemblyHands.

--pair determines which dual-camera pair is used for adaptation for evaluation. Possible values: 0,1 0,2 0,3 1,2 1,3 2,3.

--gpus gpu id.

We also provide other arguments for adjusting the hyperparameters in our S2DHand architecture, which could be found in our paper.

For example, run code like:

python3 adapt_detnet_dual.py -trs ah -tes ah --root_idx 0 --pic 1024 --resume -eid 37 --epochs 10 --start_epoch 1 --gpus 0 --checkpoint in_dataset_adapt --setup 0 --pair 1,2

If first running this code, the pretrain/ckp_detnet_37.pth model will be adapted to the camera pair 1,2 of headset 0. 1024 images pairs will be used and there are 10 adaptation epochs in total.

Test

All arguments are the same as training.

Usually, extra testing is not necessary, as it will be automatically done after adaptation. Still, we can run code like:

python3 adapt_detnet_dual.py -trs ah -tes ah --evaluate -eid 37 --gpus 0 --pic -1 --checkpoint pretrain --setup 0 --pair 1,3

This will test the pre-trained weights pretrain/ckp_detnet_37.pth on the camera pair 1,3 of headset 0.

Note: the result printed is NOT the final result on the specific dual-camera pair. It only contains SINGLE-VIEW evaluation results.

Running calc_metrics.py is needed to get two metrics on the pair we adapt to. These two metrics are the final results, which are described in our paper.

python3 calc_metrics.py --checkpoint in_dataset_adapt/evaluation/ah --setup 0 --pair 1,2

The --checkpoint, --setup, --pair arguments share the same definition with adaptation and testing. The result should be like:

1-setx-x,x: num:..., Mono-M:..., Dual-M:..., ...
2-setx-x,x: ...
...
10-setx-x,x: ...

Please refer to run.sh for a complete procedure from training to testing.

In-dataset & Cross-dataset adaptation

Setting -eid=37 indicates in-dataset adaptation, while -eid=68 indicate cross-dataset. This is because ckp_detnet_37.pth is trained on AssemblyHands while ckp_detnet_68.pth is trained on RHD and GANerated Hand.

Remember to set --root_idx=0 for in-dataset adaptation while --root_idx=9 for cross-dataset.

Visualization

Run visualize.py for visualization. However, cropped assemblyhands images are required. Please refer here for our cropped images.

Please run the following command to register the cropped data for visualization.

ln -s assemblyhands_crop data

Then, run code like:

python3 visualize.py --checkpoint in_dataset_adapt/evaluation/ah -eid 10 --sample_idx 0 --setup 0 --pair 1,2

Note: the visualization requires log file, which is generated by testing. Therefore, it is necessary to run testing (python3 train_detnet_dual.py -trs ah -tes ah --evaluate) before visualization. You may want to set --pic as a small value for a shorter testing time.

The --sample_idx indicates the index of sample for visualization. If it goes well, we can see the predictions and ground-truth labels on the images pair we choose.

Citation

If this work or code is helpful in your research, please cite:

@inproceedings{liu2024single,
 title = {Single-to-Dual-View Adaptation for Egocentric 3D Hand Pose Estimation},
 author = {Liu, Ruicong and Ohkawa, Takehiko and Zhang, Mingfang and Sato, Yoichi},
 booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
 pages = {0--0},
 year = {2024}
}

If you are using the detnet network in this project, please also cite its original paper:

@inproceedings{zhou2020monocular,
  title={Monocular real-time hand shape and motion capture using multi-modal data},
  author={Zhou, Yuxiao and Habermann, Marc and Xu, Weipeng and Habibie, Ikhsanul and Theobalt, Christian and Xu, Feng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5346--5355},
  year={2020}
}

This project is built on an unofficial PyTorch reimplementation of Zhou et al. Monocular real-time hand shape and motion capture using multi-modal data. CVPR2020. Original project page: https://github.com/MengHao666/Minimal-Hand-pytorch.

Contact

For any questions, including algorithms and datasets, feel free to contact me by email: lruicong(at)iis.u-tokyo.ac.jp