register new input for a model

[bobluoluo]

In addition to the keypoint training config for coco dataset, I added 2 new fields for each person annotation, a face box field, a facial landmark field. Landmarks follows the convention of keypoint field x,y,v. So my person data looks like the following right now: annotation: {bbox:[], mask:[], keypoints:[], #and extra: face_box[], face_keypoints:[]}

I have created a data_mapper and registered the extras into the object instance. so now an instance would contain gt_box, gt_class, gt_mask, gt_keypoints, gt_face_box, gt_face_keypoints.

So, my question is what is the fastest way to registering these extras for the input of a keypoint detection model. How to I register loss functions, etc and the default one only supports these. Please point me an example or something. Thanks ahead.

Does this repo have a paper of explaining how everything works?