search

NVIDIA / DIGITS

Deep Learning GPU Training System
https://developer.nvidia.com/digits
BSD 3-Clause "New" or "Revised" License
4.12k stars 1.38k forks source link

Calculating class-wise accuracy of Semantic segmentation #1632

Closed akbwaj closed 7 years ago

akbwaj commented 7 years ago

I have been able to perform transfer learning on FCN-8s for a 2-class (background,foreground) problem in DIGITS and the results look good on my test set. For calculating area under intersection of foreground class, I need to generate these images. I see that DIGITS is overlaying a binary-segment on original image and dispalying the segmented images. I was expecting the binary-segment to be like feature labels with a pixel value of 0 (background) or 1 (foreground). But the pixel labels are in a range. What is the threshold that I should use in this case to produce the same results as I see when DIGITS makes it?

akbwaj commented 7 years ago

It was only an issue of the result being in RGB format, figured it out. Thanks.