Hi, thank you for releasing Mars-Bench and the mb-mars_seg_msl dataset.
We are currently using mb-mars_seg_msl for semantic segmentation experiments and noticed a possible issue in the validation split.
For some validation samples, the image and the corresponding segmentation mask do not appear to match. In particular, the spatial structures and terrain/object regions in the mask seem inconsistent with those in the image.
We have checked our data loading and preprocessing pipeline, and the mismatch still exists when directly visualizing the original image-mask pairs from the validation set.
Could you please confirm:
Whether there is any known issue with the image-mask pairing in the validation split?
Whether there is a specific indexing or mapping rule that should be used to correctly associate the images and masks?
Whether an updated or corrected version of the validation split is available?
Thanks for your help and for making this benchmark publicly available.
Hi, thank you for releasing Mars-Bench and the mb-mars_seg_msl dataset.
We are currently using mb-mars_seg_msl for semantic segmentation experiments and noticed a possible issue in the validation split.
For some validation samples, the image and the corresponding segmentation mask do not appear to match. In particular, the spatial structures and terrain/object regions in the mask seem inconsistent with those in the image.
We have checked our data loading and preprocessing pipeline, and the mismatch still exists when directly visualizing the original image-mask pairs from the validation set.
Could you please confirm:
Whether there is any known issue with the image-mask pairing in the validation split?
Whether there is a specific indexing or mapping rule that should be used to correctly associate the images and masks?
Whether an updated or corrected version of the validation split is available?
Thanks for your help and for making this benchmark publicly available.