REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation

REBASE addresses training-free in-context segmentation, where a single annotated reference image introduces a new object category at inference time. The paper argues that shared backgrounds between reference and query images create spurious cross-image similarities and weaken prompt localization. It identifies a low-rank background feature subspace from the reference image, then projects both reference and query features onto its orthogonal complement in closed form. Positive point prompts are generated with similarity-weighted farthest-point sampling and a refined dense similarity prior. The abstract reports state-of-the-art results among training-free methods on PACO-Part, FSS-1000, and cross-domain datasets including ISIC2018.