Weight Magnitude

[Apeke]

Hi and thanks for the paper about BNF

I am very interested in the edge detector. You have used in your paper the Weight Magnitudes, but I didn't found the explanation, how they are calculated or where they can be found. Can you give me a hint, please?

Thanks a lot André

[gberta]

Hi André,

Our globalization method can be used with many different edge detectors. I didn't make our edge detector code available since there are more powerful edge detectors available out there now. The description of how I computed the weights used for boundary detection can be found in the original BNF paper in section 3.2.1. Note, that you can also use our globalization code without the boundaries at all (just using the RGB affinities), which works reasonably well too.

[Apeke]

Hi Gedas,

I was confused about the Figure 3 and the Weight Magnitude, because I thought you get the Weight Magnitude directly from the feature maps and the network itself. Of course, I have read 3.2.1 before, but now I understand that you used the computation from 3.2.1 also for Fig. 3.

I do not need the RGB affinities, because in my project, we do not have any RGB picture. That is why I try to rebuild the edge detector and want to use boundary affinities. Your edge detector looks very useful, because he has strong semantic boundaries around the sought object.

Thanks for your answer.

[kohjingyu]

@gberta When I run the demo BNF_multi_class_edge_affinity_demo.m for the train image, (2008_000003.png), the output appears slightly different from the paper. Is this due to implementation differences in the code published?

Thanks, Jing Yu

[gberta]

This is probably due to implementational differences. The paper and the github versions are slightly different.