Shrink boxes only visible to front-camera in nuScenes

[xryang22]

Hi, thank you for your great work which I found still inspiring among today's domain adaption algorithms. I have a question after looking thoroughly into your code. In https://github.com/cxy1997/3D_adapt_auto_driving/blob/master/convert/nusc2kitti.py#L114 it is specified the camera name as 'CAM_FRONT', as kitti only support one camera, and in https://github.com/cxy1997/3D_adapt_auto_driving/blob/master/convert/nusc2kitti.py#L333 the boxes unvisible to the front camera is ignored. Hence, in the fine-tune process after stat norm, the model just ignore other boxes? Is that reasonable?

Thank you, looking forward to your reply.

[xryang22]

In fact, I found this issue when I'm checking the gt label number of one specific sample. The sample token is "e106cd7b9867481a8280b2a8cc9ff5a5" with 64 gt boxes in nuScenes, while in the label_2/003635.txt in kitti format of this sample, there are only 12 lines, debugger shows that there are 50 boxes are not visible to the front camera, so they are discarded during the conversion from nuscenes to kitti, and the left two belong to non-vehicle classes are ignored as well.