Evaluation of Monocular 3D Object Detection

[skmhrk1209]

Hello,

In my understanding, KITTI-360's evaluation protocol for 3D object detection is designed for LiDAR-based 3D object detectors that take an accumulated point cloud as an input and are not designed for monocular 3D object detectors. To evaluate monocular 3D object detectors, is it reasonable to merge the predicted bounding boxes from all the frames in the window of each accumulated point cloud, via 3D/BEV NMS and perform the evaluation code? Otherwise, is there a better way? Thank you!

[yiyiliao]

Hi, thank you for your interest in our leaderboard. Yes, the current leaderboard is not designed for monocular 3D object detectors. It is quite reasonable to merge the monocular-based predictions for evaluation on our leaderboard.

[skmhrk1209]

OK. Thanks!