yezhen17 / 3DIoUMatch

[CVPR 2021] PyTorch implementation of 3DIoUMatch: Leveraging IoU Prediction for Semi-Supervised 3D Object Detection.
155 stars 16 forks source link

Compared with sess #2

Closed czy341181 closed 3 years ago

czy341181 commented 3 years ago

Congratulation! a good paper!

I want to ask the difference between SESS and this paper. Does SESS have no pseudo-label mechanism? But this paper first introduces the filtering module of pseudo-label mechanism, and then adds an IOU module? Is my understanding correct?

yezhen17 commented 3 years ago

Thanks, @czy341181, I think your understanding is correct.

Strictly speaking, SESS has no such mechanism, which you can check in the paper or code. In a word, SESS enforces a set of consistency losses to all predicted bounding boxes. However, you can also view the raw outputs of the teacher network in SESS as some kind of pseudo-label. Then the pseudo-labels are not standard because they are different from the format of ground truth labels. That's why I say SESS has no pseudo-label mechanism strictly speaking.

Yes, our paper introduces the filtering (and deduplication) of pseudo-labels, and then we find only filtering based on classification and objectness is not good enough because we ignore the localization confidence. As a solution, we design an IoU module.