Verified, then logged.
Sun 22 Nov 2026 · 00:16 KST / 15:16 UTC한국어 | ENSearch /

Breaking01 / AI

Image segmentation method finds look-alikes without training

The method finds candidate objects, compares them with a reference image and outlines the matching object.

A monitor displays a reference image of a speckled ceramic mug alongside several similar mugs, with an amber segmentation mask covering only the matching mug. AI-generated image

Researchers presented an image segmentation method that they said reduced matching costs by more than 100 times without additional training. Their paper, published on 13 Nov, described how the system finds the object that matches a user-provided reference image.

The paper described personalised image segmentation as identifying the area occupied by a specific object supplied by the user. The system first uses an existing detector to find candidate objects regardless of their category. It then compares the candidates with the reference using visual features from an existing model.

The paper's top-K Sinkhorn method compares only the image patches that match most confidently. Its cost comparison uses partial optimal transport, a method that optimises matches between selected portions of two inputs.

The researchers attributed earlier methods' problems to missing the correct candidate rather than imprecise object outlines. Their system uses SAM, an image segmentation model, to refine the selected candidate's outline.

The researchers reported comparisons with PerSAM, P²SAM and PDM on a new evaluation dataset, alongside PerSeg, PerMIS and DAVIS. These evaluations were reported by the researchers in the paper.

The paper's abstract did not disclose the size or detailed composition of the new evaluation dataset.