The num of boxes of matching_gt_bbox is more than that of valid_gt? about lidar_rcnn HOT 6 CLOSED

@QingXIA233
Hi, qing
We use an old-school detection label assignment, assigning a gt box to every positive box. And we judged the positive and negative samples in the loader.
There is a one-to-one assignment in DETR, which I guess is your want.

from lidar_rcnn.

Ok, got it, your method makes sense to me now. Thank you for the quick reply. I'll do it according to your method. Perhaps I'll also perform some experiments with label assignment similar to DETR later.

from lidar_rcnn.

HI，I am quite confuse about 'gt_cls = 0', I checked the code in 'get_matching_by_iou', it assign the first gt_box to the proposal who match none with all the gt_boxes and coresponding 'gt_cls=0'. Same as this way it did in other places suce as 'relabel_by_iou'. Question is, 'gt_cls=0' is`n vehicle in waymo dataset?(not background class) @Lzc6996

from lidar_rcnn.

@qingzhouzhen
HI, gt_cls = 0 is background.

from lidar_rcnn.

@qingzhouzhen HI, gt_cls = 0 is background.

If I transfer dataset to others such as nuscenes, will thie logic work probably？

from lidar_rcnn.

@qingzhouzhen Hi, in my understanding, in this repo, they start the class labels from 1 instead of 0, which means 0 becomes the label that is not considered. My own data' label starts from 0(Car), so I add 1 to each of them before I feed the data to LiDAR R-CNN, then I minus 1 when I run create_results.py after testing the model.

from lidar_rcnn.