The results on test set are very similar with MonoDETR

[pummi823]

Dear author: on car val set(AP_3D), your method exceeds MonoDETR by 1.3~2% on three difficulty, but on test set, yours only exceed MonoDETR by around 0.5%, why the gap is so huge?

[Senwang98]

@pummi823
Ha ha ha, funny issue.
You have found the key to the problem.
Some metrics of ADD are even inferior to MonoDETR ( Hard 3D AP), which is why the author does not add MonoDETR in "Table 2".
BTW, the author seems to have abandoned the open source plan of ADD.

[Senwang98]

@pummi823
Back to the question itself, on the one hand, the number of train sets is small, and the training error is relatively large; while the test set is relatively large, so the training error is very small. On the other hand, if you use a student network with poor performance, the benefits are very obvious, but if you use a student network with strong performance, the performance improvement is not necessarily so obvious, and it does not necessarily increase strictly linearly.

[jichaofeng]

@rockywind @pummi823 @Senwang98
We have read the paper,but there are some question for the teacher model. If we use the GT depth for teacher model, does the depth estimation network in teacher model still need training ? When we use monodetr as baseline method, the feature distillation is applied to which features ? Thank you very much.

[yexiguafuqihao]

We even don't known how to train the teacher model. This paper is not qualified to be accepted as an AAAI paper.

[Senwang98]

@yexiguafuqihao I agree with you, the performance is too poor.

[HHuiwen]

We even don't known how to train the teacher model. This paper is not qualified to be accepted as an AAAI paper.

News! This paper has even been accepted by TPAMI: 'Advancing 3D Object Detection with Depth-Aware Spatial Knowledge Distillation.'