Can't find concentration with a truncated loss mentioned in paper (Section:Moving as a unit.Eq.(6))

[Hevans123]

Hi, thank you for providing your project. In this code, I can't find concentration regularization with a truncated loss in region tracking process mentioned in paper (Section:Moving as a unit.Eq.(6)). Could you please tell me where it is? Or if you did'nt upload the part of concentration regularization with a truncated loss, could you please upload this part?

[sunshineatnoon]

Hi, the concentration loss is a truncated loss. Eq. (6) means that we only compute concentration loss within a patch (within a h*w region), thus we first use im2col to convert the feature map to a set of patches and compute concentration loss within each patch. Please let me know if this makes sense.

[Hevans123]

Hi, thank you for your reply. But it still confuses me. It seems that the concentration loss is following this section, which did'nt use truncation operation.

And I can't find the truncation operation mentioned in Eq.(6) in the code, why is concentration loss introduced with a truncated operation in the paper?

[Zhongdao]

Hi, great work!
I am also a little confused about the truncation operation.
It seems the implementation be like: only pixels within the patch are used for computing the loss.
But the description in the paper is like: only pixels outside the patch are used for computing the loss.
Maybe I am wrong, please kindly let me know if so. Appreciate any reply!

[Zhongdao]

I think both ways make sense but their objectives are different