Auto-detect segmentation layers

[fm3]

When cubing datasets the user has to specify which layers are segmentation layers. It would be helpful to add a heuristic that auto-determines this. E.g. find patches of uniform color/id vs noisy images (which are probably raw data). It may be possible to run this heuristic only on a sampled subset of the data

[philippotto]

A very cheap heuristic could be the following:

• sample N bboxes à 4x4x4 pixel (ignore bounding boxes which are all zero)
• per bbox, count the unique values
• compare average unique values per bounding box against a threshold (for segmentation, I would assume the value of unique ids to be <= 2)

[jstriebel]

Dataset.from_images currently has a cheap heuristic by checking if segmentation or labels is a substring in the relative path of the layer. A data-driven heuristic would be even better.