Awesome

Identity-Aware Hand Mesh Estimation and Personalization from RGB Images

[ECCV2022] "Identity-Aware Hand Mesh Estimation and Personalization from RGB Images".

Introduction

This repo is the PyTorch implementation of ECCV2022 paper "Identity-Aware Hand Mesh Estimation and Personalization from RGB Images". You can find this paper from paper link.

Install

Environment for IDentity-Aware Hand Mesh Estimation (IdHandMesh)

conda create -n IdHandMesh python=3.8   
conda activate IdHandMesh

Please follow official suggestions to install pytorch and torchvision. We use pytorch=1.7.1, torchvision=0.8.2
Requirements

Please see environment.yaml.
Install MANO from MANO.
Note that torch_sparse and MPI-IS Mesh are required for the CMR model (one of the methods our method is compared to).

a) If you have difficulty in installing torch_sparse etc., please use whl file from here.

b) MPI-IS Mesh: We suggest to install this library from the source.

More details can be found at CMR.

Dataset

Dex_YCB

Please download DEX-YCB dataset from this link, and create a soft link in data, i.e., data/dex_ycb. Inside the folder 'datasets', run the following scripts to process the data.
```
python split_annotations_with_cropping.py
python get_rest_hand_mesh.py
```

After processing the data, you should have the following directory struction.

${ROOT} 
|-- conv
|   |-- ...
|-- data
|   |-- dex_ycb
|   |   |-- 20200709-subject-01
|   |   |-- 20200813-subject-02
|   |   |-- ...
|   |   |-- 20201022-subject-10
|   |   |-- 20201022-subject-10
|   |   |-- 20201022-subject-10
|   |   |-- bop
|   |   |-- calibration
|   |   |-- cropped_hand_size_2         # generated by our .py file
|   |   |-- models
|   |   |-- split_annotations           # generated by our .py file
|-- datasets
|   |-- dex_ycb
|   |   |-- ...
|-- options
|-- out
|-- scripts
|-- src
|-- template
|   |-- dex_ycb_j_regressor.npy
|   |-- MANO_RIGHT.pkl
|   |-- template.ply 
|   |-- transform.pkl
|-- utils
|   |-- ...
|-- ....py

Reproducing the baseline and our method.

Training

a) Train the baseline model.
```
./scripts/train_dex_ycb_mano_based_baseline.sh
```
b) Train the our model with ground truth hand shape.
```
./scripts/train_dex_ycb_mano_based_our_model_with_gt_shape.sh
```
c) Train baseline with confidence branch. (only train the confidence branch, other parts are frozen).
```
./scripts/train_dex_ycb_mano_based_conf_branch.sh
```

Run hand shape calibration.

a) Get results from the baseline model

python mis_dex_ycb_get_predictions_baseline_with_conf.py

b) Perform calibration

python calibrate_from_shape_params.py

Evaluate the performance without optimizatin module.

a) Baseline performance
```
./scripts/eval_dex_ycb_mano_based_baseline.sh
```
b) Our model when fed with groundtruth hand shape
```
./scripts/eval_dex_ycb_ours_gt_hand_shape.sh
```
c) Our model when fed with calibrated hand shape
```
./scripts/eval_dex_ycb_ours_calibrated_hand_shape.sh
```
Optimizatin module during inference.

Get 2d predictions.
```
python mis_dex_ycb_get_predictions_2d.py
```
Run optimization and evaluate at the same time.

a) Baseline performance
```
python optimization_dex_ycb_baseline.py
```
b) Our model when fed with groundtruth hand shape
```
python optimization_dex_ycb_ours_with_gt_hand_shape.py
```
c) Our model when fed with calibrated hand shape
```
python optimization_dex_ycb_ours_with_calibrated_hand_shape.py
```

Reproducing the methods being compared to.

CMR

./scripts/train_dex_ycb_cmr.sh
./scripts/eval_dex_ycb_cmr.sh

Boukhayma’s model

./scripts/train_dex_ycb_boukhayma
./scripts/eval_dex_ycb_boukhayma.sh

Metro

Install Metro and then use the following script to generate the tsv file for DEX_YCB.
```
metro_utils/tsv_demo_dex_ycb.py
```

Reference

@inproceedings{kong2022identity,
title={Identity-Aware Hand Mesh Estimation and Personalization from RGB Images},
author={Kong, Deying and Zhang, Linguang and Chen, Liangjian and Ma, Haoyu and Yan, Xiangyi and Sun, Shanlin and Liu, Xingwei and Han, Kun and Xie, Xiaohui},
booktitle={Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part V},
pages={536--553},
year={2022}
}

Acknowledgement

Our implementation is developed with the help of the following open sourced projects:

Please also consider cite the above projects if you find them helpful.