search

Cc-Hy / CMKD

Cross-Modality Knowledge Distillation Network for Monocular 3D Object Detection (ECCV 2022 Oral)
Apache License 2.0
107 stars 9 forks source link

Questions about teacher models #45

Closed RookieHuang closed 1 year ago

RookieHuang commented 1 year ago

Anyone have an idea about why CenterPoint and PointPillar are the worse teacher models than SECOND, even if they perform better than SECOND?

Cc-Hy commented 1 year ago

Hi, Sorry for being late. According to our earlier experiments, a better teacher model will not necessarily lead to better student results. And similar conclusions have been reached in other papers. I think it's more important to make it more suitable for students to learn from the teacher models than to make the teacher models themselves better. For example, I have attempted to use some two-stage teachers, which are much better in performance. However, the results are not better. In many cases, the improvement in the performance of the lidar model comes from its deeper extraction of point cloud information, which may be difficult for the student model to learn. Hope the above can help you a little.

RookieHuang commented 1 year ago

Thank for your reply!