Awesome

Aligning Motion Generation with Human Perceptions

This repository contains the PyTorch implementation of the paper "Aligning Motion Generation with Human Perceptions".

framework

Quick Demo

MotionCritic is capable of scoring a single motion with just a few lines of code.

bash prepare/prepare_smpl.sh

from lib.model.load_critic import load_critic
from parsedata import into_critic
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
critic_model = load_critic("critic/motioncritic_pre.pth", device)
example = torch.load("criexample.pth", map_location=device)
example_motion = example['motion'] # [bs, 25, 6, frame], rot6d with 24 SMPL joints and 1 XYZ root location
# motion pre-processing
preprocessed_motion = into_critic(example['motion']) # [bs, frame, 25, 3], axis-angle with 24 SMPL joints and 1 XYZ root location
# critic score
critic_scores = critic_model.module.batch_critic(preprocessed_motion)
print(f"critic scores are {critic_scores}") # Critic score being 4.1297 in this case

https://github.com/ou524u/AlignHP/assets/92263178/edd9600a-5c72-4594-80b3-356d442736c9

Try scoring multiple motions with some more code

bash prepare/prepare_demo.sh

from lib.model.load_critic import load_critic
from render.render import render_multi
from parsedata import into_critic
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
critic_model = load_critic("critic/motioncritic_pre.pth", device)
example = torch.load("visexample.pth", map_location=device)
# calculate critic score
critic_scores = critic_model.module.batch_critic(into_critic(example['motion']))
print(f"critic scores are {critic_scores}")
# rendering
render_multi(example['motion'], device, example['comment'], example['path'])

https://github.com/ou524u/AlignHP/assets/92263178/a11fa74d-43a4-4dff-a755-c0c9fe00ccfe

Getting Started

Setup the Environment

conda env create -f environment.yml
conda activate mocritic

Task Documentation

Dataset & Pretrained Model

Download the pre-processed datasets and pretrained models:

bash prepare/prepare_dataset.sh  # Download pre-processed datasets
bash prepare/prepare_pretrained.sh  # Download pretrained models

Alternatively, you can manually download the files from the following links:

Pre-processed datasets: Google Drive Link
Pretrained MotionCritic model: Google Drive Link

Build Your Own Dataset (Optional)

To build your own dataset from the original motion files and annotation results:

bash prepare/prepare_fullannotation.sh
bash prepare/prepare_fullmotion.sh

Manual downloads are available here:

Full annotation results: Google Drive Link
Complete motion .npz files: Google Drive Link

After pre-processing the complete data, build your dataset with:

cd MotionCritic
python parsedata.py

Evaluating the Critic Model

Reproduce the results from the paper by running:

cd MotionCritic/metric
python metrics.py
python critic_score.py

Train Your Critic Model

Train your own critic model with the following command:

cd MotionCritic
python train.py --gpu_indices 0 --exp_name my_experiment --dataset mdmfull_shuffle --save_latest --lr_decay --big_model

Critic Model Supervised Fine-Tuning

First, prepare the MDM baseline:

bash prepare/prepare_MDM_dataset.sh
bash prepare/prepare_MDM_pretrained.sh

If you encounter any issues, refer to the MDM baseline setup.

Next, start MotionCritic-supervised fine-tuning:

cd MDMCritic

python -m train.tune_mdm \
--dataset humanact12 --cond_mask_prob 0 --lambda_rcxyz 1 --lambda_vel 1 --lambda_fc 1 \
--resume_checkpoint ./save/humanact12/model000350000.pt \
--reward_model_path ./reward/motioncritic_pre.pth \
--device 0 \
--num_steps 1200 \
--save_interval 100 \
--reward_scale 1e-4 --kl_scale 5e-2 --random_reward_loss \
--ddim_sampling \
--eval_during_training \
--sample_when_eval \
--batch_size 64 --lr 1e-5 \
--denoise_lower 700 --denoise_upper 900 \
--use_kl_loss \
--save_dir save/finetuned/my_experiment \
--wandb my_experiment

Additional Python scripts for various fine-tuning purposes can be found in MDMCritic/train, detailed in the fine-tuning documentation.

Citation

If you find our work useful for your project, please consider citing the paper:

@article{motionpercept2024,
    title={Aligning Motion Generation with Human Perceptions},
    author={Wang, Haoru and Zhu, Wentao and Miao, Luyi and Xu, Yishu and Gao, Feng and Tian, Qi and Wang, Yizhou},
    year={2024}
}

Acknowledgement

If you use MotionCritic in your work, please also cite the original datasets and methods on which our work is based.

MDM:

@inproceedings{
  tevet2023human,
  title={Human Motion Diffusion Model},
  author={Guy Tevet and Sigal Raab and Brian Gordon and Yoni Shafir and Daniel Cohen-or and Amit Haim Bermano},
  booktitle={The Eleventh International Conference on Learning Representations },
  year={2023},
  url={https://openreview.net/forum?id=SJ1kSyO2jwu}
}

HumanAct12:

@inproceedings{guo2020action2motion,
  title={Action2motion: Conditioned generation of 3d human motions},
  author={Guo, Chuan and Zuo, Xinxin and Wang, Sen and Zou, Shihao and Sun, Qingyao and Deng, Annan and Gong, Minglun and Cheng, Li},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={2021--2029},
  year={2020}
}

FLAME:

@inproceedings{kim2023flame,
  title={Flame: Free-form language-based motion synthesis \& editing},
  author={Kim, Jihoon and Kim, Jiseob and Choi, Sungjoon},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={37},
  number={7},
  pages={8255--8263},
  year={2023}
}

UESTC:

@inproceedings{ji2018large,
  title={A large-scale RGB-D database for arbitrary-view human action recognition},
  author={Ji, Yanli and Xu, Feixiang and Yang, Yang and Shen, Fumin and Shen, Heng Tao and Zheng, Wei-Shi},
  booktitle={Proceedings of the 26th ACM international Conference on Multimedia},
  pages={1510--1518},
  year={2018}
}

DSTFormer:

@inproceedings{zhu2023motionbert,
  title={Motionbert: A unified perspective on learning human motion representations},
  author={Zhu, Wentao and Ma, Xiaoxuan and Liu, Zhaoyang and Liu, Libin and Wu, Wayne and Wang, Yizhou},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={15085--15099},
  year={2023}
}

SMPL:

@incollection{loper2023smpl,
  title={SMPL: A skinned multi-person linear model},
  author={Loper, Matthew and Mahmood, Naureen and Romero, Javier and Pons-Moll, Gerard and Black, Michael J},
  booktitle={Seminal Graphics Papers: Pushing the Boundaries, Volume 2},
  pages={851--866},
  year={2023}
}

We also recommend exploring other motion metrics, including PoseNDF, NPSS, NDMS, MoBERT, and PFC. You can also check out a survey of different motion generation metrics.