Segmentation

Per-pixel semantic segmentation on a frozen-encoder token grid. The default path upsamples the grid with a neural decoder; the rows below are the substrate and component choices a segmentation run can swap in, with the walkthrough for each where one exists.

Method	Summary	Walkthrough
Neural decoder (default)	A lightweight conv decoder regresses a per-pixel class map; the SegmentationHead trains it with cross-entropy + soft-Dice.	Dense prediction
Decoder-free pixel classifier	Swap the neural decoder for a per-pixel classifier on the encoder’s own per-head attention (XGBoost / RF / logistic / MLP) — no decoder, no checkpoints.	Attention-based segmentation
Multi-encoder composite	Concatenate the dense outputs of several foundation models into one richer per-position vector (the paper’s +7.95% mean Dice headline).	Composite

See also

The dense contract, config, metric, and outputs are on the Segmentation reference.