search

TUI-NICR / ESANet

ESANet: Efficient RGB-D Semantic Segmentation for Indoor Scene Analysis
Other
236 stars 51 forks source link

Void pixels in ground truth labels #31

Open ShanyCohen opened 3 years ago

ShanyCohen commented 3 years ago

Hello,

I'm trying to figure out how the void pixels affect the mIou and the loss computation. The labels are being decremented by 1 in the code so that class numbers in the prediction and the ground truth labels will match. So to my understanding, there are pixels with the value -1 in the label image but the prediction doesn't include void pixels. Are those pixels being ignored or being considered as true/false predictions?

Thanks, Shany

mona0809 commented 3 years ago

Void is ignored during both training and mIoU. For cross entropy loss the index -1 is ignored (see: https://github.com/TUI-NICR/ESANet/blob/main/src/utils.py#L27-L30) For mIoU, both prediction and ground truth are filtered, such that only those pixels are taken into account, that are not void in ground truth. (see: https://github.com/TUI-NICR/ESANet/blob/main/train.py#L437-L460)