This is a pure pytorch re-implementation of the SCRFD face detector by insightface original repo

The metrics provided by the current repo are slightly different due to the difference in post-processing. The models are tested to output the same raw result.

• The original implementation of SCRFD is based on mmdetection framework which is not quite flexible
• The pytorch-only implementation is a way easier to integrate into your own pipeline avoiding digging into a mmdet dependency hell
• The current version provides a more clean and easy-to-understand code compared to mmdet implementation
• The Mmdet implementation is not TorchScript-friendly

• Converted weights Vanilla to Current (the conversion script is attached)
• Reimplemented SCRFD models
• Refactored Loss computation
• Improved BboxSafeRandomCrop augmentation proposed in the original SCRFD paper, implemented with the albumentations library
• TorchScript conversion and inference example
• Onnx conversion and inference example
• Evaluation results matrix

tbd