New EASe framework improves unsupervised segmentation by capturing fine-grained details in complex scenes using foundation models.

EASe introduces Semantic-Aware Upsampling with Channel Excitation (SAUCE) to enhance low-resolution feature representations from foundation models

Addresses limitations of existing patch-level approaches that lose fine-grained structural details needed for complex multi-component morphologies

Operates as a domain-agnostic unsupervised segmentation method, requiring no labeled training data across different image types

Combines feature calibration and self-supervised upsampling techniques to improve mask discovery accuracy in challenging real-world scenes