train-DeepLab icon indicating copy to clipboard operation
train-DeepLab copied to clipboard

Pascal VOC mIOU should ignore `void` labels

Open bermanmaxim opened this issue 8 years ago • 2 comments

Hello, At training, the void label in Pascal VOC is mapped to the background class 0. However, for evaluation purposes, it would make sense to ignore pixels corresponding void pixels in the ground-truth, as this is standard for Pascal VOC evaluation. If I'm not mistaken, this step was not done here. I opened a similar issue on this tensorflow implementation.

bermanmaxim avatar Jan 19 '17 12:01 bermanmaxim

@bermanmaxim I really find this problem. How do you update the mIOU code?

Hutai avatar Mar 24 '17 02:03 Hutai

You can just zero the prediction mask at the locations where the void pixels are before computing the mIoU.

IssamLaradji avatar Jul 18 '18 23:07 IssamLaradji