COB
COB copied to clipboard
kmaninis
about the HED and CEDN scores
In the paper Convolutional Oriented Boundaries, you offered the hed and cedn scores(ODS) on PASCAL Context:
If it's convenient, could you offer your retrained model of these two method？or the predicted-images?（cause I cant get the result you presented in your paper） it will help me a lot，thank you！
Hi, you can download the images one by one, using wget, eg for CEDN:
wget https://data.vision.ee.ethz.ch/kmaninis/share/COB/Results/CEDN/2008_000003.png
or
wget https://data.vision.ee.ethz.ch/kmaninis/share/COB/Results/HED_trainval/2008_000003.png
for HED.
I would have sent you the pre-computed results in a zip file if I had access to the files, but I don't anymore.
thank you！
so you train CEDN and HED in this dataset to get these scores?
but in CEDN paper ,the author use more images to train and less images to test(val)
and get lower score(57 in the paper), am I miss some important things? What is the essential difference between these two datasets that I don't see（Because I don't think that subtle differences can lead to a big difference in scores）
its confused me a lot.
and if convinent，could you tell me how to make the object-counter(edges-GT) from Semantic segmentation？for example，when i use pascal2012's Semantic segmentation to get the object-counter, i found there are some white area:
How should we remove these white areas and keep only the edges of objects？
and another question is how to make these thick edges become thin? actually i have some object counter from the paper:Object Contour and Edge Detection with RefineContourNet
but i really dont know how to get the final GT:
if it's convinent, could you offer your code for making edge-GT? or juet offer the images that have already processed.
thank you again, it will help me a lot!
We did re-train the models using the same images in the comparisons.
As for the second question, we used MATLAB's bwmorph(grad, 'thin', Inf), where grad is the gradient magnitude of the image, if I remember correctly.