charigyang / motiongrouping

Repository accompanying the paper: Self-supervised Video Object Segmentation by Motion Grouping. C. Yang, H. Lamdouar, E. Lu, A. Zisserman, W. Xie.
BSD 2-Clause "Simplified" License
240 stars 35 forks source link

Training in supervised learning manner #6

Closed fupiao1998 closed 3 years ago

fupiao1998 commented 3 years ago

Hello, thank you for such valuable work. I am curious about the effect of such a structure under the framework of supervised learning. In other words, analyze in principle whether the current framework is suitable for training under supervision. Looking forward to your answer, thank you!

charigyang commented 3 years ago

Hi, we expect it to work better than self-supervised for sure given GT supervision. i.e. train on the masks directly instead of our proxy reconstruction loss. We're not sure by how much quantitatively, as it's not what we're trying to show in this work. Although many work has already done it the supervised way with a different architecture, and those are probably more optimised for supervised setting, if that is what you are looking for.

fupiao1998 commented 3 years ago

Ok, I get it, thanks for such a quick reply.