MMDetection's implementation of BEVFusion fails when input_modality = dict[lidar = False, camera = True]. Trying to understand why.

[abubake]

I am trying to get the mmdetection3d bevfusion project to train on nuscenes dataset using camera only. I managed to run inference using a converted version of mit's camera-only-detection pretrained model but am unable to get it to train. I think I am confused about the actual functionality of the input modality dict.

Should I actually be setting lidar = False when trying to train camera only? Or is this actually unrelated somehow? trying to make sense of what my issue is. Currently when I set lidar = False I get an error that 'n008-2018-08-01-15-52-19-0400__LIDAR_TOP__1533153553948088.pcd.bin' filepath doesn't exist.

Any points in the right direction would be greatly appreciated :) TIA

[abubake]

Solution: You most definitely need to set lidar=false, and there are several classes used for processing the lidar and point data that need to be commented out/ not used in the config for it to train.