xuebinqin
commented
1 year ago This is a good question. Since other task now is dichotomous image segmentation, we didn't define the to-be-segmented targets explicitly. The to-be-segmented targets are actually defined by our dataset. As for your example of the bike and people, we would like to segment both human and bike. Because in practical applications like image editing, users are more likely to segment the foreground which in your example includes both human and bike. Another example in our dataset is the table_and_chair category. Because table and chair sometimes appear close to each other and are hard to be excluded. Hence, both are segmented in some scenarios. In addition to defining the categories, we also need to think about (or guess) what are expected by different applications and users. According to our experiences, when a human is riding a bike, users are more likely to segment both. In our dataset collection process, we take the above concerns into consideration. We don't follow the previous rules which define the task by the semantic meaning or characteristic of the targets. The key rule followed by our data collecting and annotation excluding any conflicting annotations, which is the basis of our task. Because deep models are actually fitting a transformation between input and output. It really doesn't care about if you are segmenting human, satellite image, or medical image. If your model capacity is enough, you can even combine all the dichotomous image segmentation tasks together. The only premise is that your annotations of the dataset are non-conflicting. So that is why our task is defined as Dichotomous Image Segmentation other than certain object segmentation. Following this rule, our DIS task is an open framework and can be further expanded by adding more non-conflicting samples, which is currently ongoing for our DIS 2.0.
On Sun, Jul 17, 2022 at 11:04 PM xianglei3 @.***> wrote:
I am a student of Tsinghua University. I want to ask whether u2net only supports single category segmentation. For example, for the picture of people riding on bicycles, can people and bicycles be separated separately? Thank you. Thank you very much
— Reply to this email directly, view it on GitHub https://github.com/xuebinqin/U-2-Net/issues/316, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORNPG5EPULC7ZPF7PWLVUTXWLANCNFSM533BZKFA . You are receiving this because you are subscribed to this thread.Message ID: @.***>
-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/
I am a student of Tsinghua University. I want to ask whether u2net only supports single category segmentation. For example, for the picture of people riding on bicycles, can people and bicycles be separated separately? Thank you. Thank you very much