Awesome

Motion Matters: Neural Motion Transfer for Better Camera Physiological Sensing

:book: Abstract

Machine learning models for camera-based physiological measurement can have weak generalization due to a lack of representative training data. Body motion is one of the most significant sources of noise when attempting to recover the subtle cardiac pulse from a video. We explore motion transfer as a form of data augmentation to introduce motion variation while preserving physiological changes of interest. We adapt a neural video synthesis approach to augment videos for the task of remote photoplethysmography (rPPG) and study the effects of motion augmentation with respect to 1. the magnitude and 2. the type of motion. After training on motion-augmented versions of publicly available datasets, we demonstrate a 47% improvement over existing inter-dataset results using various state-of-the-art methods on the PURE dataset. We also present inter-dataset results on five benchmark datasets to show improvements of up to 79% using TS-CAN, a neural rPPG estimation method. Our findings illustrate the usefulness of motion transfer as a data augmentation technique for improving the generalization of models for camera-based physiological sensing.

:wrench: Setup

STEP1: bash setup.sh

STEP2: conda activate ma-rppg-video-toolbox

STEP3: pip install -r requirements.txt

STEP 4: Install PyTorch using the below command,

pip install torch==1.8.2 torchvision==0.9.2 torchaudio==0.8.2 --extra-index-url https://download.pytorch.org/whl/lts/1.8/cu111

The exact versioning may vary depending on your computing environment and what GPUs you have access to.

STEP 5: Download the appropriate pre-trained model and place it in the appropriate folder within checkpoints/. We recommend using Vox-256-New over Vox-256-Beta. You will also have to place the appropriate .yaml file in the config/ folder, and make sure it's appropriately pointed to using the --config parameter when running augment_videos.py.

:computer: Usage

This motion-augmentation pipeline supports three rPPG video datasets - UBFC-rPPG, PURE, and SCAMPS. The pipeline utilizes an unofficial pytorch implementation of the paper One-Shot Free-View Neural Talking-Head Synthesis for Video Conferencing by Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu (NVIDIA). A link to the original, unofficial implementation can be found here. Python 3.6.13 and Pytorch 1.8.2 are used.

Below is a basic example of utilizing augment_videos.py to augment all subjects provided in the UBFC-rPPG with motion based on a supplied directory of driving videos and the Vox-256-New pre-trained model:

python augment_videos.py --config ./checkpoints/checkpoint_new/vox-256-spade.yaml --checkpoint ./checkpoints/checkpoint_new/00000189-checkpoint.pth.tar --source_path /path/to/source/dataset/folder --driving_path /path/to/driving/dataset/folder --augmented_path /path/to/augmented/dataset/folder/to/generate --relative --adapt_scale --dataset UBFC-rPPG

Note that an augmented output path is specified with --augmented_path. The augmented output includes the exact same folder structure as the input dataset (e.g., UBFC-rPPG dataset's DATASET2 folder structure), and contains all of the corresponding ground truth files that are copied over. More information about supported datasets, as well as a Google Drive link to videos that we used as driving videos for our experiments in the paper, can be found in the datasets section below. Please also note that the generated videos, especially when stored as .npy files, are quite large (e.g., ~5GB per motion-augmented UBFC-rPPG video) due to being stored without any kind of additional compression.

Multiprocessing can be enabled with the --mp command line option to speed-up the dataset generation process. Depending on your computing environment, this may be the best way to generate augmented datasets using this GitHub repo. It is strongly recommended that you utilize CUDA_VISIBLE_DEVICES alongside the aforementioned python command when using multiprocessing (e.g., CUDA_VISIBLE_DEVICES=0,1,2,3 python augment_videos.py ...) in order to limit the number of GPUs available for multiprocessing, which in turn will influence how many processes are spawned. For reference, using 4 x NVIDIA RTX A4500 GPUs it takes approximately 22 hours to augment all 42 videos from the UBFC-rPPG dataset with motion.

You may find code for motion analysis and visualization using OpenFace to be useful here, in the rPPG-Toolbox repo. We also provide pre-trained models in pretrained_models/ for use with the rPPG-Toolbox. You can train with your own, motion-augmented data by following the instructions in the rPPG-Toolbox repo here.

:sparkles: Pretrained Models

The below pre-trained models were obtained from here.

A Google Drive back-up of these pre-trained models (in the form of a checkpoints/ folder that can be directly used in this repo) can be found here.

:file_folder: Datasets

The following source rPPG video datasets can be downloaded via the below links (note that aside from SCAMPS, these datasets require some sort of interaction with the authors and/or a filled out form):

• UBFC-rPPG
• PURE
• SCAMPS

The TalkingHead-1KH driving videos we used, filtered using either the mean standard deviation in head pose rotations or the mean standard deviation in facial AUs as described in our paper, can be found here. In addition to citing our work if you use this motion augmentation pipeline, please cite the TalkingHead-1KH dataset if you use the aforementioned driving videos. Our self-captured, constrained driving video set (CDVS) as described in the paper will be released soon.

:scroll: Acknowledgments

Thanks to zhanglonghao1992, NV, AliaksandrSiarohin, and DeepHeadPose for their useful code implementations!

:scroll: Citation

If you find our paper or this toolbox useful for your research, please cite our work.

@inproceedings{paruchuri2024motion,
  title={Motion Matters: Neural Motion Transfer for Better Camera Physiological Measurement},
  author={Paruchuri, Akshay and Liu, Xin and Pan, Yulu and Patel, Shwetak and McDuff, Daniel and Sengupta, Soumyadip},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={5933--5942},
  year={2024}
}