We compare different foundation models (SAM and DINO variants) with classical features in ilastik and fully supervised learning. The FM features outperform classical approaches. Attentive probing yields better performance than RF, but is too slow for interactive use. See example results in the fig.
