search

LittlePey / SFD

Sparse Fuse Dense: Towards High Quality 3D Detection with Depth Completion (CVPR 2022, Oral)
Apache License 2.0
264 stars 35 forks source link

giou-->diou #21

Open CBY-9527 opened 1 year ago

CBY-9527 commented 1 year ago

Hi, I have a question. I replaced giou in the configuration file with diou, and the code can be trained normally without encountering the Nan situation, which may be because diou considers the situation where RoI and gt do not intersect. Generally speaking, the effect of diou should be better than that of giou. However, after I replaced giou with diou, the detection performance decreased by 1 to 2 percentage points.

LittlePey commented 1 year ago

Hi, we have not tried training SFD with DIoU. We didn't encounter Nan situation when training SFD with GIoU, maybe you need change an environment.

Ultra11nstinct commented 1 year ago

@CBY-9527 Hi, sry to bother, but did you solve the NaN problem in training stage? I have encounter the same problem.