Verify that the post-processing merged into FaceMesh works correctly. The object detection model can be anything other than BlazeFace. YOLOv4 and FaceMesh committed to this repository have modified post-processing. - View it on GitHub