search

yezhen17 / 3DIoUMatch

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

Error occurred when remove_empty_box. #4

Closed czy341181 closed 3 years ago

czy341181 commented 3 years ago

In the inference phase, there is a step before nms that needs to remove 128 proposals with few points,

but when I get the data like 001333, because this data is special like a face and there is no gt in the whole scene. causing all my boxes to be removed in this step, thus causing errors. Have you met a similar error?

d14bd1535b1dcee044a1106881014c6 301cba4ffec7e9b82f2896f84d97fea e196aaf9ac3f042f57ee3936e4b9bc8

czy341181 commented 3 years ago

Removing this remove_empty_box module directly should not have much effect on the results, right?

yezhen17 commented 3 years ago

Hi @czy341181 ,

I have not met with this problem because I always set remove_empty_box in config_dict to be False here, which is by default. I believe you can simply ignore this, because VoteNet also sets this to False.

czy341181 commented 3 years ago

Ok, Thanks for your timely reply!
By the way, I noticed that sess only trained 120 epochs, but 3DIoUmatching trained 1001, is this because IoU is hard to train?

yezhen17 commented 3 years ago

Well, I don't think it's because IoU is hard to train. It seems to me that the pseudo labels need a longer training schedule to reach a high quality, and then the student can effectively learn from the pseudo labels.

By the way, we also tried training SESS with the same schedule as ours, but the performance did not exceed the original performance.

czy341181 commented 3 years ago

Yes, I also try setting the threshold on SESS to select high quality proposal, the same training time is not as good as before. So, a longer training schedule will improve performance?

yezhen17 commented 3 years ago

Yes, I also try setting the threshold on SESS to select high quality proposal, the same training time is not as good as before. So, a longer training schedule will improve performance?

I'm not sure about how will SESS behave after setting a threshold, but I think you can take a try to lengthen the training schedule.

czy341181 commented 3 years ago

Thanks , I have no problem