Segmentation only resulting one label

[ramdhan1989]

hi,
thank you for expanding the code for multiple images. I tried using my own datasets containing thin section of rock images where image size is various. However, the training run smoothly and stop at very early epoch due to the labels reach min number of labels.
do you have a suggestion on how to improve the result? the segmentation result contains only single color for every grid.

thank you
Regards,

[giacomopiccinini]

Hey, thanks for reaching out!

I have admittedly not maintained this repo in a while, but we could probably figure that out.

I would start from setting an higher value for --min_classes. This should control when the training stops: that is, if you reach the minimum number of classes, the script is exited. It seems this might be your case? From what I gather from your message, you get each image segmented with just one colour/region. Is that right?

You could also set an higher value form --n_classes: this parameter regulates how many different regions the images are started with.

If after experimenting with these two parameters you still encounter issues, could you please provide e.g. (part of) dataset so that I can test it myself?

I reckon another possibility is that the multi-image extension is not 100% right. I could probably look into that, too.

Please let me know if you can make any progress!

[ramdhan1989]

Hi,
Yes you are correct the output only give me one color and nothing segmented. If I use one image the result still make sense similar to original repo. I tried to increase --n_classes but it didn't help.

thank you