PJLab-ADG
Question on generalization study on Uni3D
In Table 8 of Uni3D, generalization study is conducted by evaluating zero-shot detection accuracy on KITTI. Since two separate detection heads are trained and a dual-BN is leveraged for nuscenes and waymo during pre-training, what are the implementation details for conducting zero-shot detection on KITTI and which detection head is used?
@wangyubin04
Hi, thanks for your question.
Note that Table 8 in Uni3D reports the AP_{BEV} and AP_{3D} of the car category at IoU = 0.7. The experiment setting follows that of ST3D, where the baseline model is trained on source domain (e.g., Waymo) and tested on target domain (e.g. KITTI). For example, the first and fourth rows represent the results of the PV-RCNN model pre-trained on Waymo but evaluated on KITTI (only car category). Similarly, we can use Uni3D to perform the joint training on Waymo and nuScenes, and then directly test on the KITTI dataset, which is shown in the third row of Table 8.
