Loss function: additional robustness?

[blukaz]

Hi @yuxumin, I've been wondering if you tried to combine both the CD_l1 & CD_l2 loss function during training? Maybe it would yield better results?

[yuxumin]

Hi, @blukaz . We haven't tried it.

[blukaz]

I'll try it out and let you know.
Btw, have you tried EMD? Do you know any good and valid implementation for that?

[yuxumin]

@blukaz Thanks a lot. We haven't tried to optimize emd loss along with CD loss during the training, but i think it worth a try. As for implementation for emd, maybe you can try it (But I can't guarantee the correstness of this implementation)

[blukaz]

@ thank you for the link! Keep you updated in the coming days.

[blukaz]

@yuxumin do you have any recommendation on how to fix clustering of the coarse prediction? I have the problem that the dense loss is a magnitude lower than the coarse. Did you encounter similar problems?

[yuxumin]

You mean that the CD-loss for dense cloud pairs are much less than those for coarse pairs? Yes, i notice it too. I find some works adopt the weighted sum of these two losses. For example, loss = fine_loss + \alpha * coarse_loss and decrease the value of \alpha during the training. But i am not sure whether it will help.

[blukaz]

Yes, exactly that. Ok, I see. I'll try that. Thank you for your quick response!