Question about feature-based KD

[Senwang98]

@monster-ghost
I found that feature-based KD using the whole foreground area, but in your paper, you use Gaussian-like mask.
Is the whole foreground have better than Gaussian-like mask?

[monster-ghost]

Firstly，we use the whole foreground area in the feature-based KD and gaussian-like mask in the result-based KD. And
Secondly, I think it is better to use the whole foreground area in the feature-based KD. However, more precise areas are necessary to avoid noise in the result-based KD. So we employ gaussian-like mask in result-based KD.

[Senwang98]

@monster-ghost
Hello, I would like to know: Was the final result of MonoDistill on the kitti list supervised by the teacher ckpt(trained by "train" split) you provided?

If we want to submit the kitti, have you retrained the teacher (based on trainval)?

Will this bring performance improvement?