Comments on Waymo Validation Results?

[jyhjinghwang]

Hi,

I'm looking into the monocular 3D object detection on Waymo Open Dataset.
In your paper, you showed results in this table:

In the recent CVPR paper, "MonoJSG: Joint Semantic and Geometric Cost Volume for Monocular 3D Object Detection", they summarized the results in this table:

Their test set results on KITTI seem to be better than CaDDN.

All in all, your CaDDN achieves 5% AP at 0.7 IoU whereas MonoJSG is barely 1% AP at the same threshold. I believe both of your methods used the same setting, i.e., front cameras and vehicle class.

Do you have any ideas for where the difference might come from?
Thanks a lot!