Brute-Force Facial Landmark Analysis

[arXiv]

Dependency

• VLFeat
• MatConvNet (tested with commit d62881db)

Usage

1. Download the pre-trained model and extract to models/
2. Run Test.m

Face detection

The detection for the example images are provided. However, to run on new images, a face detector is required. We recommend using MTCNNv2 due to its robustness and stability. Also, our detection refinement module is trained with MTCNNv2 using its default parameters.

The accepted format of the bounding box is [x y width height] (no need to round to integer), different from the output of the detect_face function in MTCNNv2. It can be transformed using the following code:

bbx(:, 3:4) = bbx(:, 3:4) - bbx(:, 1:2);

Videos

• Temporal smoothing under complete occlusion (provided detection)
• Interactive conditional prediction - eye corner
• Interactive conditional prediction - nose tip

Citation

If you use this code for your research, please cite the paper:

@article{BFFL2018,
  title={Brute-Force Facial Landmark Analysis With A 140,000-Way Classifier},
  author={Li, Mengtian and Jeni, Laszlo and Ramanan, Deva},
  journal={AAAI},
  year={2018}
}