Dear organizers,
Two questions:
- Will geodata for the validation / test set images be released so that we can join different layers to these images? This would provide additional value to the SEN12MS dataset as high-resolution ground truth exists for these sets!
- Are there plans to release the methods by which the high-resolution ground truth labels were collected (or just the human labeled points)? Accuracy of the GT labels would allow us to reason about the best possible performance we can hope for from a model.
Best,
Caleb
Dear Caleb,
the answer to both questions is YES. We are currently removing some minor errors in the land cover labels of the land cover channels NOT intended for use in the competition (i.e. the LCCS land cover maps). Once this is done, we intend to publish the full DFC2020 dataset including disclosure of the geographical regions. Furthermore, there will be the final JSTARS paper summarizing the competition, which will also include some information about the quality of the reference data.
Best regards,
Michael