My question is basically about the codes in 581 lines of detector3d_template.py
rcnn_recalled = (iou3d_rcnn.max(dim=0)[0] > cur_thresh).sum().item()recall_dict['rcnn_%s' % str(cur_thresh)] += rcnn_recalled
Here the max value in iou3d are used to count the how many predict box can pass the corresponding iou thresholds.
However, for evaluation, the final_box was selected by the cls_score, instead of the IOU between the ground truth and itself.
So this will actually lead to a miss-match for the output, we use the IOU of the best-match proposal to calculate the performance instead of the box selected by cls_score.
I want to know where I went wrong, thanks for any help!
My question is basically about the codes in 581 lines of detector3d_template.py
rcnn_recalled = (iou3d_rcnn.max(dim=0)[0] > cur_thresh).sum().item()recall_dict['rcnn_%s' % str(cur_thresh)] += rcnn_recalledHere the max value in iou3d are used to count the how many predict box can pass the corresponding iou thresholds. However, for evaluation, the final_box was selected by the cls_score, instead of the IOU between the ground truth and itself. So this will actually lead to a miss-match for the output, we use the IOU of the best-match proposal to calculate the performance instead of the box selected by cls_score.I want to know where I went wrong, thanks for any help!