facexlib Documentation

facexlib provides ready-to-use face-related functions based on SOTA open-source methods. The library includes the following functional modules:

• Detection: Face detection (Source: Pytorch_Retinaface)
• Alignment: Face alignment (Source: AdaptiveWingLoss)
• Recognition: Face recognition (Source: InsightFace_Pytorch)
• Parsing: Face parsing (Source: face-parsing.PyTorch)
• Matting: Face matting (Source: MODNet)
• Headpose: Head pose estimation (Source: deep-head-pose)
• Tracking: Face tracking (Source: SORT)
• Assessment: Image quality assessment (Source: hyperIQA)
• Utils: Face restoration helpers

FaceXLib supports three distinct facial landmark detection modes, providing varying levels of granularity for facial alignment and analysis:

• 5 landmarks: Typically used for basic facial orientation and alignment (eyes, nose, and mouth corners).
• 68 landmarks: A standard dense landmark set used for detailed facial shape and expression analysis.
• 98 landmarks: An even denser landmark set for high-precision facial modeling.

Install the facexlib package using pip. Ensure you have Python >= 3.7 and PyTorch >= 1.7 installed. Using Anaconda or Miniconda is recommended.

pip install facexlib

FaceXLib automatically downloads pre-trained models during the first inference. If you have an unstable network connection, you can manually download the models and place them in the following directory:

PACKAGE_ROOT_PATH/facexlib/weights