search

Megvii-BaseDetection / BEVDepth

Official code for BEVDepth.
MIT License
688 stars 96 forks source link

Intuition on how depth correction works #115

Closed Francis777 closed 1 year ago

Francis777 commented 1 year ago

Hello BEVDepth authors! After reading your paper, I highly agree with the motivation of depth correction, but I don't really get how it would work:

If $D^{gt}$ is incorrect because of the incorrect lidar-camera calibration, and instead of adjusting that directly, we make our DepthNet to fit better to this incorrect $D^{gt}$(by increasing the receptive field as the paper mentioned), isn't the depth prediction $D^{pred}$ still mis-aligned with $F^{2d}$? In other words, $D^{pred}$ isn't aligned with the original image anymore but $F^{2d}$ is.

Maybe I'm missing something obvious, look forward to your reply!

yinchimaoliang commented 1 year ago

Hello, Thanks for your attention. The purpose depth refinement module is to align the gap between image feature and depth generated by lidar. By doing that, it is beneficial for the training process of depth. During inference time, it is also beneficial for unifying all camera depth to one coordinate(lidar).

Francis777 commented 1 year ago

Thanks for your clarification!