question about the version of Waymo Datase

[seonhoon1002]

Thank you for sharing your great work.

I have a question about Waymo Dataset

Table 2 in your paper, It was measured by Waymo dataset. But I wonder what is the version you used.

Summary: what did you use version of Waymo?

[XuyangBai]

Hi, I use v1.2

[seonhoon1002]

There are some weird points, because performances of PVRCNN and LiDAR RCNN in table 2 are too low.

Bellow are performance measured by v1.2 in SST[1] and PV-RCNN++[2] Vehicle(reported in SST) Pedestrian(reported in SST) (reported in PV-RCNN++)

reference [1] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, Zhaoxiang Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 8458-8468 [2] Shi, Shaoshuai, et al. "PV-RCNN++: Point-voxel feature set abstraction with local vector representation for 3D object detection." arXiv preprint arXiv:2102.00463 (2021)

[XuyangBai]

The results of PVRCNN are borrowed from CenterPoint, and the results of LiDAR-RCNN are taken from the LiDAR-RCNN paper

The table here is the results for multi-class detection while the ones you provide is one model trained on one specific class (I assume) ?

[seonhoon1002]

Q1: The table here is the results for multi-class detection while the ones you provide is one model trained on one specific class (I assume) ?

A1: Vehcle in SST, yes you're right(but pedestrain trained on multi-class), I shouldn't bring vehicle result in SST. I think LiDAR-RCNN result is reasonable.

But in PV-RCNN case, I think there is a problem in Centerpoint. Because they report Waymo V1.0 result not V1.2(V1.0 is lower than v1.2). Please refer bellow table(refer from PV-RCNN original paper[1])

I think It is easily confused if not informed version of Waymo in paper.

I think you should edit pv-rcnn result if you can.

reference [1] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3D object detection. CVPR, 2020. 1

[XuyangBai]

Oh thanks for pointing out this information which I did not notice before, I will edit this result if I plan to have a new version.