Fast-BEV with CenterPoint head
Hello,
I am trying to use M2BEV / Fast-BEV or at least the m2BEV projection step for image features to BEV. This does seem to work and I do get everything in the correct shapes, but I am limited to using the CenterPoint detection head. Now my training loss is not decreasing after the first steps and I am wondering if its possible to use the CenterPoint head with your architecture or if a learned assignment of detections -> GT is required as in your head?
Same problem, have you resolved it yet?
Its been a while and to be honest I do not remember if I was able to solve it. However, conceptionally it would be problematic to use the centerpoint head, as the features are scattered across the complete ray. So it would have no indication at which point the object should be located.
However, what is fed into the centerpoint head is a (C, BEVH, BEVW) feature, so will it still be affected by rays?
Have you up the project ? can you provide me steps to build and run properly ?