nms error about fast-rcnn HOT 5 CLOSED

Hi, I got the same error.
Could you share your solutions? Thanks.

from fast-rcnn.

It is the problem of proposal bounding box.
You should check whether your proposal bounding box's width and height bigger than 0. MCG sometimes return a non-valued bounding box. Selective search has no such problem.

Anyway, check your bounding box at first.

from fast-rcnn.

Thanks for your quick reply and good suggestion.

It is weird because the bounding box is generated by selective search in my experiments.

from fast-rcnn.

Thank for your hints @sxjzwq .

I think the bounding-boxes returned from the bounding-box regresser may not follow the "well-known" convention. In other words, the area of returned bounding-box may be 0 or negative. @rbgirshick

from fast-rcnn.

Yes. And if the area of the bounding-box is zero, there will be more problems in the training.

In fast-rcnn/lib/roi_data_layer/roidb.py, line114-line115,
targets_dx = (gt_ctr_x - ex_ctr_x) / ex_widths
targets_dy = (gt_ctr_y - ex_ctr_y) / ex_heights
if the ex_weight == 0, the bounding-box regression targets will be a extremely large number.
I simply remove those zero area bounding boxes before this step,

from fast-rcnn.