Home

Awesome

Novel-view Synthesis and Pose Estimation for Hand-Object Interaction from SparseViews

Official PyTorch implementation of the ICCV 2023 paper

Project page | Paper

Table of contents

Installation

PyTorch with CUDA support are required. Our code is tested on python 3.9, torch 1.13.0, CUDA 11.6, and RTX 3080 Ti. Clone the repository and install the dependencies:

git clone https://github.com/iscas3dv/HO-NeRF
cd honerf
conda create -n honerf python=3.9 -y
conda activate honerf
pip install torch==1.13.0+cu116 torchvision==0.14.0+cu116 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu116
pip install -r requirements.txt

We also need to install pytorch3d from here.

Preparing

Preparing Datasets

Download the HandObject dataset and name it 'data' and put it in the root directory.

Download Pre-trained Models (Optional)

Download the Pre-trained Models which is named as 'exp' and put it in the root directory.

Training and Testing

Offline stage training and testing

Run the following command to train the offline stage object and hand models:

# Object model
python exp_runner.py --mode train --conf ./confs/wmask_realobj_bean.conf --case bean --gpu 0
# Hand model
python exp_runner.py --mode train --conf ./confs/wmask_realhand_hand1.conf --case hand1 --gpu 0

The results will be saved in ./exp/(CASE_NAME)/wmask_realobj(hand)/checkpoints

Run the following command to test the offline stage models to get rendering results or mesh models:

# Object model
python exp_runner.py --mode test --conf ./confs/wmask_realobj_bean.conf --case bean --gpu 0 --is_continue
# Hand model
python exp_runner.py --mode mesh --conf ./confs/wmask_realhand_hand1.conf --case hand1 --gpu 0 --is_continue

The results will be saved in ./exp/(CASE_NAME)/wmask_realobj(hand)/test_render(meshes)

Online stage fitting

Fitting on single frame

First run the following command to perform optimization based on render loss and pose regularizer loss:

python fitting_single.py --conf ./fit_confs/fit_1_8views.conf --case 1_8view --gpu 0

The results will be saved in ./fit_res/view_8/1/obj_name/frame_name/pose_1

Then run the following command to perform optimization based on render loss, pose regularizer loss and interaction loss :

python fitting_single.py --conf ./fit_confs/fit_12_8views.conf --case 12_8view --gpu 0

The results will be saved in ./fit_res/view_8/12/obj_name/frame_name/pose_12

Fitting on video

Based on the above command, we obtain the results based on single frame optimization and save them in ./fit_res/view_8/12/.

(Optional 1)Then we run the following code to implement video-based optimization using stable loss:

python fitting_video.py --conf ./fit_confs/fit_123_8views_0.conf --case 123_8view_id0 --gpu 0

The results will be saved in ./fit_res/view_8/123/123/obj_name/frame_name/pose_4

(Optional 2)We can also run the following code to implement video-based optimization using stable loss and stable contact loss:

python fitting_video.py --conf ./fit_confs/fit_1234_8views_0.conf --case 1234_8view_id0 --gpu 0

The results will be saved in ./fit_res/view_8/1234/obj_name/frame_name/pose_4

Evaluation

Our results

You can also download the Our results named 'fit_res.zip'.

Get more results

We need to obtain the hand-object interaction reconstruction results, penetration points, and rendering results to evaluate our method. Run the following command to get the online stage results:

# Get reconstruction results for fit_type: '1'
python get_res.py --conf ./fit_confs/get_res_1.conf --case get_res_1 --gpu 0
# Get reconstruction results and penetration points for fit_type: '12'
python get_res.py --conf ./fit_confs/get_res_12.conf --case get_res_12 --gpu 0
# Get penetration points for fit_type: '123'
python get_res.py --conf ./fit_confs/get_res_123.conf --case get_res_123 --gpu 0
# Get penetration points for fit_type: '1234'
python get_res.py --conf ./fit_confs/get_res_1234.conf --case get_res_1234 --gpu 0
# Get rendering results for fit_type: '12'
python get_res.py --conf ./fit_confs/get_render_type12.conf --case render_res_type12 --gpu 0 --render True

The reconstruction results will be saved in ./fit_res/view_num/fit_type/obj_name/frame_name/mesh_*.

The penetration points will be saved in ./fit_res/view_num/fit_type/obj_name/frame_name/inner_*.

The rendering results will be saved in ./fit_res/view_num/fit_type/obj_name/frame_name/render_*.

Evaluation

Then run the following command to get the evaluation metrics.

#Get the pose error.
python ./analys_results/analys_hand_obj_pose.py
#Get the interaction error.
python ./analys_results/analys_interaction.py
#Get the acceleration error.
python ./analys_results/analys_acc_err.py
#Get the PCI error.
python ./analys_results/analys_pci.py
# Get the rendering quality error.
python ./analys_results/analys_psnr_ssim_lpips.py

Citation

@inproceedings{qu2023novel,
  title={Novel-view Synthesis and Pose Estimation for Hand-Object Interaction from Sparse Views},
  author={Qu, Wentian and Cui, Zhaopeng and Zhang, Yinda and Meng, Chenyu and Ma, Cuixia and Deng, Xiaoming and Wang, Hongan},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={15100--15111},
  year={2023}
}

This codebase adapts modules from NeuS, HALO and Pytorh3D. Please consider citing them as well.