mask-modal strategy on bevfusion

[konyul]

Thanks for your work! In paper, bevfusion can get 0.40 with mask-modal strategy when LiDAR sensor missing. But I added "ModalMask3D" to bevfusion and trained with mask-modal strategy, the result only is 0.25. Can you provide more technology details or point out my some error operations?

Originally posted by @dingmiaomiao in #93 (comment)

[curiosity654]

Hello, I'm also insterested about this problem, could you please provide more details of your experiments? E.g. which codebase do you use (mmdet3d or MIT), and the pipeline of your training (train from scratch for 20 epoch or init with pretrained for 6 epochs)? Thank you in advance.