I use my own model resulted in backbones generating 0 candidate boxes, which resulting in errors in the postprocess module

[ACFIRSTONE]

I have improved the pillars based model and will use this project for tensorrt acceleration. As only vehicles will be detected, objects cannot be detected in some scenes, so the backbone network will generate 0 candidates, which will cause the postprocess module to report an error. I added the judgment for bndbox_num in lidar-postprocess.cu before line 472 (because there is an error), and if it is 0, return, which avoids the error
if (bndbox_num==0) return;

[hehualin-tut]

I had the same question and was very lucky to have your answer, it saved me a lot of time, thanks
#@#@#@#@#@#@#
I have made a new discovery. Your approach is correct, but there is an issue. When you add a return statement at line 470, if the previous scene detected something but the current frame does not detect anything, bndbox_num_after_nms_ will not be reset to zero. It will keep returning the results from the previous detection. Therefore, the operation bndbox_num_after_nms_ = 0 needs to be placed before the return statement.