Join GitHub today
GitHub is home to over 40 million developers working together to host and review code, manage projects, and build software together.Sign up
Handle "ignore" class for semantic segmentation #783
This PR makes it so we handle an "ignore" class in a more sensible way for semantic segmentation. This establishes a convention that class 0 represents an "ignore" class of pixels that should be ignored for the sake of training and evaluation. Imagery pixels with label class 0 are now set to 0 (NODATA) to prevent the model from associating good imagery with the ignore class. Chips that only contain class 0 aren't saved since they are a waste of time to train on. And, the eval does not generate metrics for class 0.
This is tested with a unit test and also in an experiment that used class 0, where debug chips were inspected for correctness.