fdfat

Working in progess...

Fast 6DoF Face Alignment and Tracking

This project purpose is to implement Ultra lightweight 6 DoF Face Alignment and Tracking. This project is capable of realtime tracking face for mobile device.

Installation

Requirements

• torch >= 2.0
• autoalbument >= 1.3.1

Install

pip install -U fdfat

Model Zoo

TODO: add best model

Training

Prepare the dataset

This project use 3d 68 points of landmark (difference from the original 300W dataset). Please go to FaceSynthetics to download the dataset (100K one) and extract it to your disk.

Create your dataset yaml file with the following info:

base_path: <path-to-face-synthesis-dataset>/dataset_100000
train: <path-to-list-train-text-file.txt>
val: <path-to-list-val-text-file.txt>
test: <path-to-list-test-text-file.txt>

note: you can use list train file in datasets/FaceSynthetics for reference.

Start training

fdfat --data <path-to-your-dataset-yaml> --model LightWeightModel

For complete list of parameter, please folow this sample config file: fdfat/cfg/default.yaml

Validation

fdfat --task val --data <path-to-your-dataset-yaml> --model LightWeightModel

Predict

fdfat --task predict --model LightWeightModel --checkpoint <path-to-checkoint> --input <path-to-test-img>

Export

fdfat --task export --model LightWeightModel --checkpoint <path-to-checkoint> --export_format tflite

Credit

• YOLOv8 : Thanks for ultralytics awesome project, I borrow some code from here.
• Ultra-Light-Fast-Generic-Face-Detector-1MB : Thanks for your lightweight face detector
• FaceSynthetics : Thanks for expressive face landmark dataset, it's a good starting point
• head-pose-estimation : Thanks for head pose estimation code