Question about TopoFR (misalignment of paper and code implementation) ?

[Bear-kai]

I just scaned the code snippet for helping better understanding the paper. And (maybe) I found two issues:

1. The input img of compute_topological_loss() is actually the augmented version, since img_randomaug = img (link) is not a deep copy, resulting the simultaneous modifaction to img when modifing img_randomaug.
2. In GUM.py, the computation is different from Eq.(7) in the paper.

Moreover, in the Appendix A.1:

Notably, in our method, we focus on preserving the 0-dimension homology H0 in the topological structure alignment loss Lsa. Because preliminary experiments demonstrated that using the 1-dimension or higher-dimension homology only increases model’s training time without clear performance gains.

How do you experimented with H1 or higher-dimension homology? The released topology.py only supports H0.

[sunbaigui]

hi @DanJun6737 ,Please take a look at this when you get a chance.

[DanJun6737]

Hi @Bear-kai ~
Thank you for your attention and suggestions regarding our work！

A1： Thank you for pointing out this issue. We overlooked this problem during the code organization process. We will fix it in the latest update.

A2： For an in-depth analysis of the computation process in GUM, you may find valuable insights in this referenced paper [1].

Ref [1] Spherical space domain adaptation with robust pseudo-label loss. CVPR 2020.

A3： Currently, we have only released the code for aligning topological structures in 0-dimension. In the future, we will release the code for computing persistent homology in higher dimensions.

Overall, our source code is well aligned with the paper.