Hi, I would like to have a clarification on the prediction masks.
This is the information from the website:
"The predictions for a particular tile should be encoded as a TIFF with the Byte (uint8) data type, match the dimensions of the corresponding BD ORTHO image and contain values between 1 and 14 (inclusive)."
Nevertheless, the provided target masks have values between 0 and 15.
"Labeled class-reference from the UrbanAtlas 2012 database. In total, 12 land-use classes are considered, corresponding to the second level of the semantic hierarchy defined by UrbanAtlas. Original data are openly available as vector images at the European Copernicus program website and were used to create raster maps that geographically match the VHR tiles from BD ORTHO. They are provided as integer rasters with index labels (0 to 15; 8, 9 do not appear in the considered regions, 0, 15 are both treated as no information and ignored by the scoring function) of size ~2,000px x ~2,000px at a resolution of 50cm/px, namely 1 km² per tile."
Should we do something special with the labels 0 and 15 on the target masks ? Or maybe the labels 8 and 9 have to be removed ?
Thanks.
Posted by: jsensio @ Jan. 6, 2022, 7:38 p.m.Correct: The provided reference data does contain class ID 0 and 15. A bit simplified, these mean no data or clouds. Your predictions should not contain these values but only valid class IDs, i.e. [1,14]. If a pixel is labeled with 0 or 15 in the reference data of the validation / test set, it will be ignored (i.e. it does not matter what class is predicted for this pixel). However, if you predict 0 or 15 for a pixel that has a valid class ID in the reference, it will count as error.
Class IDs 8,9 are not present in the dataset, neither in the train nor in the val or test part. Consequently, your model should not predict them (feel free to do, but it will count as a mistake for sure). These IDs are merely kept for consistency with the UrbanAtlas and MiniFrance datasets.
Posted by: rhaensch @ Jan. 6, 2022, 7:56 p.m.