The waymo performance much lower than CaDDN?

[ZCMax]

It seems that your method achieves comparative performance with CaDDN on KITTI datasets.

[ZCMax]

Another question is that it seems that the KITTI-format labels converted for validation datasets include the bboxes whose depth <0. During calculating mAP, these bounding boxes are also counted as target labels? Would that influence the mAP on waymo dataset?