search

junjie18 / CMT

[ICCV 2023] Cross Modal Transformer: Towards Fast and Robust 3D Object Detection
Other
336 stars 36 forks source link

Question on Modal-mask training #92

Open yyunyu opened 11 months ago

yyunyu commented 11 months ago

Hi, @junjie18 in ModalMask3D function I change the https://github.com/junjie18/CMT/blob/d3b1d782ba791f28ae72cedcd069ff4ab2cbcb06/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py#L850 seed so that the camera + LIDAR training is done without LIDAR input and it shows that it doesn't work, can you explain why we have to have LIDAR data when we train the fusion model? (i.e. why it works when we test the model without lidar point, but in training it doesn't work) Thx.

Originally posted by @yyunyu in https://github.com/junjie18/CMT/issues/31#issuecomment-1841055776

zzy-ucas commented 2 months ago

Hi, have you try to compare camera+LiDAR training and only LiDAR training. I found that image branch does not work at all.