SCRFD bounding box scaling

Hi,

I am trying to use your pre-trained onnx scrfd_10g model. I also convert it to a .blob file using our internal library to work on edge devices. When running the model, I get scores, bboxes and kps from last layers without a problem. Also, the faces are detected, but bbox layer outputs are values between 0 and 5 or 7. I think they are somehow normalized. I can not figure it out how to transform the coordinates back to the image coordinates (the image size is 640x640). Could you help, please? The same goes for kps.

Thanks!

Start by reviewing the SCRFD 10g model's preprocessing and output-coordinate handling, then compare the ONNX outputs with the converted .blob outputs for a 640x640 image. Confirm how bounding-box and keypoint values are expected to map back to image coordinates and document the verified transformation.

Written by the indexing model from the issue text.