Year: 2026
Type: preprint
Abstract: Most segmentation systems infer object extent from dense semantic decisions, leaving spatial coherence as an output rather than a decision constraint. We introduce Boundary-First segmentation, which constructs closed-region hypotheses before support-conditioned identity arbitration. Unified Boundary Evidence Transport (UBET) implements this ordering without training, using frozen region proposals, analytic DINOv2 identity evidence and a geometric conflict certificate. Across seven natural-image, object, video and medical benchmarks, SAM2- and visual-only SAM3-based UBET achieve leading or competitive one-shot segmentation. Six independent implementations using diffusion, depth, superpixel and conventional edge sources further demonstrate the portability of the paradigm. Decision-order interventions, candidate-ranking analyses and oracle diagnostics separate contour availability from identity arbitration. These results establish closed-region hypothesis testing as a transferable basis for future trainable or self-learning segmentation systems. Code is available in the public GitHub repository: https://github.com/Carzon12138/UBET-Boundary-First-Segmentation.
Source: Research Square
Author: EnBao Zhang
Institution: Hefei University of Economics
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204929480