Hi @Xharlie
Thanks for the released code.
I have a question about this:
# min_instance = min_instance.view(gt_boxnum, candidate_num)max_instance = max_instance.view(gt_boxnum, candidate_num)
and
heuristic = max_instance + ex_coords_ratio / extra_coord_nums.unsqueeze(0) + (sorted_iou.unsqueeze(0) < iou_thresh) * 2.0 + (extra_coord_nums.unsqueeze(0) < 30) * 1.0
why not use min_instance? I think min_instance is more reasonable
Hi @Xharlie Thanks for the released code. I have a question about this:
# min_instance = min_instance.view(gt_boxnum, candidate_num)max_instance = max_instance.view(gt_boxnum, candidate_num)andheuristic = max_instance + ex_coords_ratio / extra_coord_nums.unsqueeze(0) + (sorted_iou.unsqueeze(0) < iou_thresh) * 2.0 + (extra_coord_nums.unsqueeze(0) < 30) * 1.0why not use min_instance? I think min_instance is more reasonable