REBASE removes background bias in training-free one-shot segmentation

Mantha Sai Gopal · hf · 2026-07-21

REBASE is a training-free in-context segmentation method that tackles a key failure mode in one-shot segmentation: shared backgrounds between the reference and query image distort cross-image similarity and hurt prompt localization.

The method estimates a low-rank background feature subspace from the reference image, projects reference and query features onto its orthogonal complement, and then forms positive point prompts with similarity-weighted farthest-point sampling plus a refined dense similarity prior. Without any training or parameter updates, it reports new SOTA among training-free methods on PACO-Part, FSS-1000, and cross-domain datasets such as ISIC2018.

Original post →

More from Research

Research channel →