search

NVlabs / FB-BEV

Official PyTorch implementation of FB-BEV & FB-OCC - Forward-backward view transformation for vision-centric autonomous driving perception
Other
608 stars 40 forks source link

whether FB-BEV uses BEV backbone? #6

Closed SPA-junghokim closed 11 months ago

SPA-junghokim commented 11 months ago

I would like to express my gratitude for your excellent research. I had always been curious about whether using both forward and backward projections for BEV representation would enhance its performance. I had contemplated taking this up as a research topic. I was particularly astounded by the improvement in performance when a module for depth was added during the backward projection.

One aspect I'm curious about is the experiment with ResNet-50 mentioned in your paper. From what I observed, the FLOPS are lower compared to when BEVDepth is reproduced. I anticipate that to achieve this, one cannot use the BEVBackbone + neck portion utilized in BEVDepth. Could you confirm if my assumption is correct?

Thank you.

zhiqi-li commented 11 months ago

We have lower FLOPs compared to BEVDepth due to we use smaller feature dimension. We still use BEV-Backbone, which is an important module for final performance.

SPA-junghokim commented 11 months ago

Thank you for your response! It cleared up a lot of my questions. Once again, I appreciate your research. I hope we can meet at a reputable conference in the future!