arXiv new/cross; ViT/MAE color representation analysis · P2 · 2026-07-17
Beyond Color Geometry：它和通用视觉自监督的关系在于：用人类 fuzzy color categories 检查 ViT/MAE/语言监督模型如何编码颜色
中高相关；详见方法、贡献和实验边界。
Beyond Color Geometry: Evaluating Human-Like Color Representations in Vision Models arXiv new/cross; ViT/MAE color representation analysis 原文链接
编号2607.13647优先级P2类别arXiv new/cross; ViT/MAE color representation analysis会议arXiv new/cross; ViT/MAE color representation analysis方法用人类 fuzzy color categories 检查 ViT/MAE/语言监督模型如何编码颜色来源arXiv / OpenReview
先说结论。它和通用视觉自监督的关系在于：用人类 fuzzy color categories 检查 ViT/MAE/语言监督模型如何编码颜色。 中高相关；详见方法、贡献和实验边界。
Figureure 1 · Overview of the evaluation pipelineFig. 1. Overview of the evaluation pipeline. World Color Survey / Munsell chips are rendered as flat patches and shaded spheres, mapped to COLIBRI fuzzy human color memberships, passed through ViT encoders, and evaluated by comparing model representational structure with fuzzy, geometric, and categorica color structure.这张图概括 Beyond Color Geometry 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。Figureure 5 · Agreement with graded human color structure, $\rho _ { \mathrm { f u zFig. 5. Agreement with graded human color structure, $\rho _ { \mathrm { f u z z y } } ,$ across relative ViT depth for all eleven encoders. Color denotes training regime; dashed lines indicate base variants.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Beyond Color Geometry 的方法或实验，请结合正文精读段落一起看。
核心问题
它和通用视觉自监督的关系在于：用人类 fuzzy color categories 检查 ViT/MAE/语言监督模型如何编码颜色。
方法拆解
用人类 fuzzy color categories 检查 ViT/MAE/语言监督模型如何编码颜色
