Issue of some operations with columns in fpn_roi_target.py

[vietth-bka]

There is a big issue with computing ctr_x and ctr_y in line126 and line127.

BFJDet/lib/det_oprs/fpn_roi_target.py

Lines 126 to 127 in 0194d5d

	ctr_x = all_rois[gt_assignment, 0] + 0.5 * (all_rois[gt_assignment, 2] - all_rois[gt_assignment, 0] + 1)
	ctr_y = all_rois[gt_assignment, 1] + 0.5 * (all_rois[gt_assignment, 3] - all_rois[gt_assignment, 1] + 1)

But it is clear that in the line 87, that all_rois is concatenated by rpn_rois and gt_rois which are matrices added with batch_id at their first columns.

BFJDet/lib/det_oprs/fpn_roi_target.py

Line 87 in 0194d5d

all_rois = torch.cat([rpn_rois[batch_roi_inds], gt_rois], axis=0)

So I think it should be skipping the first column of all_rois:
ctr_x = all_rois[gt_assignment, 1] + 0.5 * (all_rois[gt_assignment, 3] - all_rois[gt_assignment, 1] + 1)
ctr_y = all_rois[gt_assignment, 2] + 0.5 * (all_rois[gt_assignment, 4] - all_rois[gt_assignment, 2] + 1)

[afantideng]

Great thanks for picking out this bug! We are sorry for the negligence in our re-implementation. The issue has been fixed in this commit. Fortunately, these “centers" are used to compute “distances” between boxes in weighting the embedding loss. Although the absolute coordinates of these centers are totally wrong, distances between them follow similar trends with the right mode. So we find out that fixing this bug will keep almost the same final results.