Why is background class included in evaluation?

[nij...@gmail.com]

To my knowledge, background class is not used for evaluation in segmentation when there is a severe class imbalance or background does not indicate any meaningful semantics.

Video object segmentation whose goal is to find masks of the targets in videos is one of the examples.

I think that including background class in evaluation could make it unclear which method is appropriate for this application

since predicting background class well is more worthy for evaluation than predicting other small but meaningful classes.

Do you have any idea on this?

[agriculture-vision]

Hi,

Thank you for your question.

We follow the convention for semantic segmentation evaluation as done here for the PASCAL VOC dataset, which included the background class when calculating the mIoU.

The Agriculture-Vision Team