hynann/NRDF

[CVPR2024(Highlight)] NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors

NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors

Yannan He   Garvita Tiwari   Tolga Birdal   Jan Eric Lenssen   Gerard Pons-Moll

CVPR 2024 (Highlight)

   

News

• [2024-05-09] Code and pretrained models are released!

Install

The steps are described in docs/installation.md.

Download pre-trained models

The pre-trained NRDF model can be obtained from the link (amass_softplus...)

You can extract the downloaded NRDF model into a manually created folder checkpoints within the project directory

Folder structure:

NRDF
├── lib
├── examples
├── checkpoints
│   ├── amass_softplus_l1_0.0001_10000_dist0.5_eik0.0_man0.1
│   │   ├── checkpoints
│   │   │   ├── checkpoint_epoch_best.tar
│   │   ├── config.yaml
├── ...

Data downloading and processing

The steps are described in docs/data.md.

Training

Use the following command to train your own model:

python lib/core/train_nrdf.py --config=configs/train.yaml

Modify data_dir, amass_dir and experiment/root_dir with your own paths before training.

Pose denoising

Use the following command to denoise noisy poses:

python lib/exp/pose_den.py --noisy_data_path <YOUR_DATA_PATH> --model_dir <MODEL_DIR> --bm_path <SMPL_MODEL_DIR> --step_size 0.01 --iterations 200 --device cuda

Notes:

• noisy_data_path: path to the input noisy poses, default: examples/noisy_pose.npz
• model_dir: directory of the pre-trained NRDF model, default: checkpoints/amass_softplus_l1_0.0001_10000_dist0.5_eik0.0_man0.1
• bm_path: directory of the SMPL model
• Results will be saved in outputs/

Diverse pose generation

Use the following command to generate diverse poses by using NRDF:

python lib/exp/sample_poses.py --num_samples 500 --model_dir <MODEL_DIR> --step_size 0.01 --iterations 200 --device cuda

Notes:

• model_dir: directory of the pre-trained NRDF model, default: checkpoints/amass_softplus_l1_0.0001_10000_dist0.5_eik0.0_man0.1
• Results will be saved in outputs/

Inverse kinematics from partial observations

Use the following command to solve IK from partial markers:

python lib/exp/partial_ik.py --input_path <YOUR_INPUT_PATH> --model_dir <MODEL_DIR> --bm_path <YOUR_SMPL_MODEL_PATH> --occ_type left_arm_occ --obs_type marker --device cuda

Notes:

• input_path: path to the given ground truth poses, default: examples/ik_pose.npz
• model_dir: directory of the pre-trained NRDF model, default: checkpoints/amass_softplus_l1_0.0001_10000_dist0.5_eik0.0_man0.1
• bm_path: directory of the SMPL model
• Results will be saved in outputs/

Monocular 3D pose estimation from images

Coming soon

Citation

Cite us if you find this repository is helpful to your project:

@inproceedings{he24nrdf,
    title = {NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors},
    author = {He, Yannan and Tiwari, Garvita and Birdal, Tolga and Lenssen, Jan Eric and Pons-Moll, Gerard},
    booktitle = {Conference on Computer Vision and Pattern Recognition ({CVPR})},
    year = {2024},
}

Acknowledgements

Here are some great resources we benefit from:

• Pose-NDF v2 implementation: neural distance field model;
• AMASS dataset;
• VPoser: inverse kinematics code frame;
• MoSh++: index of markers on the SMPL body surface;
• pytorch3d: conversions between quaternion and axis-angle;