A doubt about global loss and refine loss

[last-one]

I find the calculation of the global loss and the refine loss is different. The refine loss ignore the valid < 0.1, which doesn't generate the loss. But in the global loss when valid < 1.1, the label change to 0 as global_label, but the global_out doesn't change, which means the global loss only focus these visible points. Is my understanding correct?

[chenyilun95]

Yes. The global loss is defined to learn the visible keypoints and provide context for RefineNet for hard keypoints prediction.

[last-one]

Thanks very much. Do you try to make these keypoints, the valid of which < 1.1, don't generate loss, just like: global loss = (global - target) (valid > 1.1) as the calculation of refine loss. But now, the global loss = (global - target (valid > 1.1)). Which one is better? The current solution not only focuses on the visible points, but also wants to make these invisible and non-existing points to be zero.

[chenyilun95]

About this details I think that is not critical. I think the later ohemlike loss is more important. And the data with valid label 1 is only small part of all.

[last-one]

Thanks for your response.