REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation
REBASE tries to make one-shot segmentation less distracted by matching background context.
The paper says shared backgrounds between a reference image and query image can inflate similarity scores outside the target object. REBASE removes a low-rank background feature subspace from both images, then uses the cleaned-up match signal to place SAM-style point prompts. It reports state-of-the-art results among training-free methods on PACO-Part, FSS-1000, and ISIC2018 without training or parameter updates. HF Daily Papers' note
The paper says shared backgrounds between a reference image and query image can inflate similarity scores outside the target object. REBASE removes a low-rank background feature subspace from both images, then uses the cleaned-up match signal to place SAM-style point prompts. It reports state-of-the-art results among training-free methods on PACO-Part, FSS-1000, and ISIC2018 without training or parameter updates. HF Daily Papers' note
score 4