about the background categories and Instances and Tasks
Never-Walk-Away opened this issue · 2 comments
Hello, thank you very much for your work. In EQL, the cross entropy loss is used in the calculation of background categories. In eqlv2, how to deal with the background category.
and
How to try to understand this sentence ：“one task with positive label and C−1 tasks with negative labels are introduced by a single instance.”
look forward to your reply ！
Suppose we have C=5 categories.
In softmax loss, the classifier output c+1 channel, the extra 1 is for background. In inference, the background scores will suppress other categories' scores since we have a softmax operation.
In sigmoid loss, the classifier output c channel. Each channel is then activated by a sigmoid function and responsible for one category. So each channel needs a gt_label (0, 1) in training.
Given a proposal of category j, for example j=2, we re-write the gt_label to [0, 0, 1, 0, 0]. That is: one instance introduces one task with positive label and C−1 tasks with negative labels.
For background task, the gt_label is [0, 0, 0, 0, 0]
I'm very sorry to reply so late. I understand this question. Thank you very much for your answer! thanks again!