MonoFaceCompute

This repository aims to facilitate preprocessing of monocular human face videos, covering a range of commonly used outputs from semantic segmentation to face tracking. The goal is to provide a convenient and coherent repository for research work.

Computations include:

• Semantic Segmentation (https://github.com/zllrunning/face-parsing.PyTorch)
• Matting (https://github.com/ZHKKKe/MODNet)
• FAN landmarks (https://github.com/1adrianb/face-alignment)
• MediaPipe landmarks (https://github.com/google-ai-edge/mediapipe)
• Face tracking (FLAME)
  • DECA, EMOCA, FaceReconstruction: https://github.com/radekd91/inferno
  • SMIRK: https://github.com/georgeretsi/smirk
• Normals estimation
  • DSINE: https://github.com/baegwangbin/DSINE
  • omnidata: https://github.com/EPFL-VILAB/omnidata
  • StableNormal: https://github.com/Stable-X/StableNormal
  • Sapiens: https://github.com/facebookresearch/sapiens.git
• Albedo estimation
  • IntrinsicAnything: https://github.com/zju3dv/IntrinsicAnything.git

Pull requests for other computations are welcome!

Setup

1. Pull the submodules: ./pull_submodules.sh
2. Run the setup script: ./setup.sh to build a conda environment with all required dependencies.
3. Download pretrained models and other required files: ./download_all_assets.sh
4. Configure your dataset according to the examples in datasets.

One dataset consists of one or multiple monocular videos. Several parameters can be tweaked, such as the strategy for cropping the videos, what face tracker to use, what dimensions the crops should be resized to or what steps of the preprocessing pipeline to run.

This was tested on Ubuntu 22.04 with a NVIDIA A5000 GPU.

Usage

All computations are aggregated in a single entry point. Run the following command to process one dataset:

python process.py --dataset datasets/example.yaml

By default, the script will run the following steps:

• Video extraction using FFMPEG
• Face detection and cropping
• Matting
• Semantic segmentation
• Landmarks detection
• Tracking
• Tracking refinement through a landmarks-based optimization

Dataset config

Supported fields of the dataset configuration files:

Parameter	Help
base_dir	Base directory from which to retrieve the video(s).
output_dir	Where to save the processed data.
shape_sequence	Name of the sequence to use for estimating face shape.
crop_mode	fixed
crop_scale	Scaling factor for the detected face boxes for cropping.
resize	What size to resize the cropped image.
smooth_tracking	Apply a low-pass filter to the optimized pose and expression values.
tracker	What face tracker to use (DECA / EMOCA / FaceReconstruction / SMIRK).
shape_tracker	Optionally specify a different face tracker for recovering shape parameters (DECA / EMOCA / FaceReconstruction / SMIRK).
steps	What steps to launch (extract, crop, matte, segment, landmarks, track, optimize, normals)
sequences	Array of: source: input video file, relative to base_dir (e.g. "1.mp4") crop_mode: fixed / constant / smooth face_selection_strategy: strategy to use for selecting a detection when there are multiple (max_confidence / leftmost / rightmost) (only used if crop_mode=constant or crop_mode=smooth) fixed_crop: [center_x, center_y, size] (only used if crop_mode=fixed)
normals_estimator	What method to use for predicting normal maps (supported: omnidata / dsine / stablenormal / sapiens). Note that some additional setup is needed for some (e.g. setup_sapiens.sh).
normals_subsample	How much to subsample the input video for predicting normal maps (e.g. 2 means compute normals for every other frame). Default: 1. Not supported for all normals estimators.
albedo_estimator	What method to use for predicting albedo maps (supported: intrinsic_anything). Note that some additional setup is needed for some (e.g. setup_intrinsic_anything.sh).
albedo_subsample	How much to subsample the input video for predicting albedo maps (e.g. 2 means compute albedo for every other frame). Default: 1.

License

We refer to the individual submodules for their licensing information.
MonoFaceCompute itself is provided under a Attribution-NonCommercial-ShareAlike 4.0 license.