search

JingchunCheng / SegFlow

Demo code of the paper "SegFlow: Joint Learning for Video Object Segmentation and Optical Flow", in ICCV 2017
178 stars 41 forks source link

Is the only one object considered? #5

Closed jakeoung closed 7 years ago

jakeoung commented 7 years ago

In DAVIS 2017 dataset, there are many objects in each image. I wondered that how we can deal with the different number of objects in cross entropy loss. But it seems that in this work, only foreground and background is considered. Is it right?

wasidennis commented 7 years ago

Yes, we deal with foreground/background segmentation in this work on DAVIS 2016 dataset. If you would like to consider different object instances, you can try these papers in http://davischallenge.org/challenge2017/publications.html, including ours with the code at https://github.com/JingchunCheng/Seg-with-SPN.

In terms of the loss for multiple objects, one way we tried before is to set (N+1) channels in cross entropy loss for N objects plus the background in a video for training. Hope this will answer your question.