search

Junjue-Wang / LoveDA

[NeurIPS 2021] LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation
344 stars 49 forks source link

Difference between 'no-data' and 'background' classes #23

Open keliive opened 2 years ago

keliive commented 2 years ago

Hi, I was wondering what is the difference between no-data and background classes in terms of their meaning (mentioned here: https://github.com/Junjue-Wang/LoveDA#dataset-and-contest), assuming that background means land-cover classes that are not of interest?

In terms of the class indices, I believe there might be an offset by 1 based on my own exploration and the class indices given here https://codalab.lisn.upsaclay.fr/competitions/421#learn_the_details-evaluation, resulting in no-data class having index -1, background 0, ..., agriculture 6.

Junjue-Wang commented 2 years ago

Yes. The no-data region shows totally (0 ，0，0) which often locates in the boundary of the image due to the geographical registration.