search

yezhen17 / 3DIoUMatch-PVRCNN

[CVPR 2021] PyTorch implementation of 3DIoUMatch: Leveraging IoU Prediction for Semi-Supervised 3D Object Detection.
51 stars 9 forks source link

Unlabeled Weight:0 #18

Closed dof-pikes closed 2 years ago

dof-pikes commented 2 years ago

Hi! Yezhen Cong:

I tested the model which train with 0 unlabeled weight. The result indicated that the model perform better than 1 unlabeled weight.

Does 0 unlabeled weight means completely learn from labeled data ?

Thanks! I'm confused for a long time.

yezhen17 commented 2 years ago

Sorry for the late reply and thanks for your interest. Yes, 0 unlabeled weight means not supervised by unlabeled data. This could happen because the dataset is very small and the performance fluctuates.

dof-pikes commented 2 years ago

Thanks for your reply！ I tried 50% setting, 0 better than 1.

And I visualized the pseudo box which are low quality.

yezhen17 commented 2 years ago

Well using 50% data, the difference of fully supervised learning should be very small. Adding supervision from unlabeled data may be harmful instead. With a low data ratio where labels are scarce, leveraging pseudo labels from the large amount of unlabeled data would be more helpful.

dof-pikes commented 2 years ago

Thanks! Emm..Maybe low quality pseudo labels are harmful? Could you illustrate why you do not apply IoU optimization ? "the features fed to the IoU estimation branch are not differentiable w.r.t. the bounding box size." Because the input of PV-RCNN not seed points?

yezhen17 commented 2 years ago

Hi @dof-pikes ,

Sorry did not see this comment. Maybe pseudo labels are harmful when there are plenty of true labels as pseudo labels have lower quality. And yes, we find PV-RCNN RoI Grid pooling not differentiable w.r.t bounding box size.