TomCYF

we find that you do the anchor matching in feature level in _match_gt_anchor(gt, anchor) Why this? Is there any mystery in training Retinaface?