search

xavysp / DexiNed

DexiNed: Dense EXtreme Inception Network for Edge Detection
MIT License
728 stars 156 forks source link

About Bsds and Biped #116

Open pu-1 opened 2 years ago

pu-1 commented 2 years ago

Hi, I have a doubt that Bsds edges predicted by Biped dataset training are better, but Bsds metrics are lower during evaluation. Is this because of the gt value of Bsds?

xavysp commented 2 years ago

Hi, I have a doubt that Bsds edges predicted by Biped dataset training are better, but Bsds metrics are lower during evaluation. Is this because of the gt value of Bsds?

In our next publication we will try to give this answer. In the meantime you can check our last publication LDC. Sorry, I didn't give you a concrete answer.

shuttle999 commented 1 year ago

I think it may be the labeling of the BSDS dataset that causes this problem, because BIPED labels all edges as much as possible, while some edges are ignored in the labeling of BSDS, causing the model trained with BIPED to detect this part of the edges, but BSDS does not have this part of the labeling, which leads to a decrease in metrics. Maybe I'm wrong, just offering a possible thought