Mismatch between paper and code

[taranalexandru]

I can see from paper that size for template's features should be 15x15x256, but because of this condition actually it is 7x7x256. Can anyone explain?

[kaizen0890]

I'm also misunderstanding why having this line on ResDownS - forward. Actually, I've tried to remove this line in order to keep output size of feature extractor are [batch, 256, 15,15] and [batch,256,31,31], then push it into RPN and Mask. The model still work as well as within this line. Btw, i also wonder that "l" and "r" are 4 and -4 respectively? Can anyone explain?

[ZacHu-ZYH]

+1

[LANMNG]

@kaizen0890 it has been explained in the paper of SiamRPN++, which is to reduce the computational burden

[LANMNG]

actually in my view, this source code is only a half-finished product

[kaizen0890]

@kaizen0890 it has been explained in the paper of SiamRPN++, which is to reduce the computational burden

Thank you so much. I think author do like that is some kind of trick in order to get high performance. Because this step will focus to the center of template object, removing some burden informations like background, noise, etc.

actually in my view, this source code is only a half-finished product

I totaly agree with you about this point. I have some experiences in re-implementing shared source code from some chinese author. Almost of tested results isn't match with their reported results in paper. Otherwise, this paper is only conference paper, not journal. This implementation is only following to preview papers like SiamFC, SiamRPN, SiamRPN++. Therefore, there are so much unclear points in this paper that the author didn't explain in paper.