FloodNet Challenge @ EARTHVISION 2021 - Track 1 Forum

Go back to competition Back to thread list Post in this thread

> IOU scores with background as a trained class

Hi,

Is it possible to obtain the results corresponding to PSPNet when its trained with background as a class? I checked the forum and found a similar question here: https://competitions.codalab.org/forums/26986/5412/

Does it increase or decrease from 79.68 - which is calculated over only 9 classes?

Thank you

Posted by: nemo31 @ May 4, 2021, 3:12 p.m.

Background is considered while calculating mIoU in the competition, but it was omitted our paper. From our experience it usually increases the mIoU, but not necessarily all the time. Hope this answers your question.

Thank you!

Posted by: binalab @ May 4, 2021, 3:41 p.m.

Some other difference between the FloodNet paper and the competition:
1. The FloodNet paper performs supervised segmentation experiments, but the competition requires semi-supervised experiments.
2. The number of images in the competition is higher than the images used in FloodNet paper.
3. FloodNet paper does not consider Background to calculate mIoU while the competition considers Background class.

Hope this helps. Thank you!

Posted by: binalab @ May 4, 2021, 4:23 p.m.

Understood. Thank you.

Posted by: nemo31 @ May 4, 2021, 6:18 p.m.
Post in this thread