arXiv new; training-free in-context segmentation · 扫读 · 2026-07-14

REBASE：它和通用视觉自监督的关系在于：用 reference-background subspace elimination 清理 VFM cross-image similarity map，偏下游但对特征匹配有启发

中相关；详见方法、贡献和实验边界。

REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation　arXiv new; training-free in-context segmentation　原文链接

编号2607.09082 优先级扫读类别arXiv new; training-free in-context segmentation 会议arXiv new; training-free in-context segmentation 方法用 reference-background subspace elimination 清理 VFM cross-image similarity map，偏下游但对特征匹配有启发来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：用 reference-background subspace elimination 清理 VFM cross-image similarity map，偏下游但对特征匹配有启发。中相关；详见方法、贡献和实验边界。

**REBASE — training-free one-shot segmentation Figure 2**REBASE — training-free one-shot segmentation Figure 2. Overview of REBASE. Given a reference image with its binary mask and a query image, both are encoded by a frozen DINOv2 backbone to extract dense patch features. The reference background subspace is then eliminated, producing a more discriminative cross-image similarity map between the support foreground and query patches. This similarity map is then converted into two complementary conditioning signals for the frozen SAM mask decoder: (i) K spatially diverse point prompts generated via similarity-weighted farthest-point sampling (SW-FPS), and (ii) a dense prior supplied to SAM’s mask-input branch. The entire pipeline is training-free, requiring no parameter updates.这张图概括 REBASE 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。

**Figure S1**Figure S1. Comparison of DINOv2-L and DINOv3-L backbones. DINOv3-L improves performance on COCO-20<sup>i</sup>, LVIS-92<sup>i</sup>, X-Ray and FSS-1000, while DINOv2-L performs slightly better on ISIC, PASCAL-Part, and PACO-Part. These results indicate that the relative strengths of the two backbones are dataset-dependent.这张图/表用于判断 REBASE 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。

核心问题

它和通用视觉自监督的关系在于：用 reference-background subspace elimination 清理 VFM cross-image similarity map，偏下游但对特征匹配有启发。

方法拆解

用 reference-background subspace elimination 清理 VFM cross-image similarity map，偏下游但对特征匹配有启发

主要贡献

中相关；详见方法、贡献和实验边界。

实验看点

实验部分建议重点看两类证据：一是作者是否把方法收益和更强数据、更长训练、更大模型区分开；二是跨模型、跨数据或跨任务迁移是否还能保留同样趋势。

局限与读法

这篇论文的结论需要结合任务设置、训练数据规模和消融实验一起看；不要只凭单个指标判断它对通用视觉表征的价值。