Question regarding build_targets about pytorch-yolov3 HOT 3 CLOSED

It is not erasing, more like for one truth bb should not correspond to two anchors boxes. So conf_mask[b, anch_ious > ignore_thres, gj, gi] = 0 initialize those possible anchor boxes to 0, and then in the later step conf_mask[b, best_n, gj, gi] = 1 assign the best iou anchor box to that truth bb.

In your example, I guess in each iteration, it is computing the anch_ious of each truth bb (for t in range(target.shape[1]). So you expect the gj and gi are different for those bbs. (under the assumption that it is not the case of multiple-label assignment). So what conf_mask[b, anch_ious > ignore_thres, gj, gi] = 0 does is not erasing.

from pytorch-yolov3.

I think either way the best iou anchor box is only assigned to one truth bb. Because if there's repeating assignment on the anchor box, the conf_mask and tcls would only be set to 1 again.

However, for the case above, it will wipe all anch_ious > ignore_thres even the ones that have another truth bb assigned.

from pytorch-yolov3.

An anchor for a given grid cell can only be assigned to one object, so that would be fine in this case, given that the specific anchor is the best match for these objects.

from pytorch-yolov3.