Skip to content
/ NeMF Public

[NeurIPS 2022] Official implementation of "NeMF: Neural Motion Fields for Kinematic Animation"

License

Notifications You must be signed in to change notification settings

c-he/NeMF

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

NeMF: Neural Motion Fields for Kinematic Animation (NeurIPS 2022)

Chengan He1, Jun Saito2, James Zachary2, Holly Rushmeier1, Yi Zhou2

1Yale University, 2Adobe Research

NeMF Overview

Prerequisites

This code was developed on Ubuntu 20.04 with Python 3.9, CUDA 11.3 and PyTorch 1.9.0.

Environment Setup

To begin with, please set up the virtual environment with Anaconda:

conda create -n nemf python=3.9
conda activate nemf
pip install -r requirements.txt

Body Model

Our code relies on SMPL as the body model. You can download our processed version from here.

Datasets

AMASS

AMASS mocap data is used to train and evaluate our model in most of the experiments. After downloading its raw data, we run the processing scripts provided by HuMoR to unify frame rates, detect contacts, and remove problematic sequences. We then split the data into training, validation and testing sets by running:

python src/datasets/amass.py amass.yaml

You can also download our processed AMASS dataset from here.

MANN

The dog mocap data in MANN is used to evaluate the reconstruction capability of our model in long sequences, whose raw data can be downloaded from here. To process the data, we first manually remove some sequences on uneven terrain and then run:

python src/datasets/animal.py dog_mocap.yaml

Our processed data can be downloaded from here.

Note: All the preprocessed data mentioned before should be extracted to the data directory. Otherwise you need to update the configuration files to point them to the path you extracted.

Quickstart

We provide a pre-trained generative NeMF model and global motion predictor. Download and extract them to the outputs directory.

Application

We deploy our trained model as a generic motion prior to sovle different motion tasks. To run the applications we showed in the paper, use:

python src/application.py --config application.yaml --task {application task} --save_path {save path}

Here we implement several applications including motion_reconstruction, latent_interpolation, motion_inbetweening, motion_renavigating, aist_inbetweening, and time_translation. For aist_inbetweening, you need to download the motion data and dance videos from AIST++ Dataset and place them under data/aist. You also need to have FFmpeg installed to process these videos.

Evaluation

To evaluate our trained model on different tasks, use:

python src/evaluation.py --config evaluation.yaml --task {evaluation task} --load_path {load path}

The tasks we implement here include ablation_study, comparison, inbetween_benchmark, super_sampling, and smoothness_test, which cover the tables and figures we showed in the paper. The quantitative results will be saved in .csv files.

To evaludate FID and Diversity, we provide a pre-trained feature extractor at here, which is essentially an auto-encoder. You can train a new one on your data by running:

python src/train.py evaluation.yaml

Train NeMF from Scratch

In our paper we proposed three different models: a single-motion NeMF that overfits specific motion sequences, a generative NeMF that learns a motion prior, and a global motion predictor that generates root translations separately. Below we describe how to train these three models from scratch.

Training Single-motion NeMF

To train the single-motion NeMF on AMASS sequences, use:

python src/train_basic.py basic.yaml

The code will obtain sequences of 32, 64, 128, 256, and 512 frames and reconstruct them at 30, 60, 120, and 240 fps.

To train the model on dog mocap sequences, use:

python src/train_basic.py basic_dog.yaml

Training Generative NeMF

To train the generative NeMF on AMASS dataset, use:

python src/train.py generative.yaml

Training Global Motion Predictor

To train the global motion predictor on AMASS dataset, use:

python src/train_gmp.py gmp.yaml

Visualization and Rendering

Our codebase outputs .npz data following the AMASS data format, thus they can be visualized directly with the SMPL-X Blender add-on. To render the skeleton animation of .bvh files, we use the rendering scripts provided in deep-motion-editing.

Acknowledgements

  • Our bvh I/O code is adapted from the work of Daniel Holden.
  • The code in src/rotations.py is adapted from PyTorch3D.
  • The code in src/datasets/amass.py is adapted from AMASS.
  • The code in src/nemf/skeleton.py is taken from deep-motion-editing and our code structure is also based on it.
  • Part of the code in sec/evaluation.py is adapted from Action2Motion.
  • Part of the code in src/utils.py is taken from HuMoR.
  • The code in src/soft_dtw_cuda.py is taken from pytorch-softdtw-cuda.

Huge thanks to these great open-source projects!

Citation

If you found this code or paper useful, please consider citing:

@inproceedings{he2022nemf,
    author = {He, Chengan and Saito, Jun and Zachary, James and Rushmeier, Holly and Zhou, Yi},
    title = {NeMF: Neural Motion Fields for Kinematic Animation},
    booktitle = {NeurIPS},
    year = {2022}
}

Contact

If you run into any problems or have questions, please create an issue or contact chengan.he@yale.edu.

About

[NeurIPS 2022] Official implementation of "NeMF: Neural Motion Fields for Kinematic Animation"

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages