arXiv new; AAAI 2026 supplementary complete version · P2 · 2026-07-21

MDND：它和通用视觉自监督的关系在于：用非可微 refinement branch 生成高质量内部目标，再监督可微 branch，无监督学习形状 correspondence

中相关；详见方法、贡献和实验边界。

MDND: Unsupervised Learning Guided by Non-Differentiable Refinement for Shape Correspondence　arXiv new; AAAI 2026 supplementary complete version　原文链接

编号2607.15887 优先级P2 类别arXiv new; AAAI 2026 supplementary complete version 会议arXiv new; AAAI 2026 supplementary complete version 方法用非可微 refinement branch 生成高质量内部目标，再监督可微 branch，无监督学习形状 correspondence 来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：用非可微 refinement branch 生成高质量内部目标，再监督可微 branch，无监督学习形状 correspondence。中相关；详见方法、贡献和实验边界。

**Figureure 2 · : A comparison of deep functional map architectures**Figure 2: A comparison of deep functional map architectures. (a) The conventional dual-branch framework which is fully differentiable and thus limited to differentiable solvers. (b) Our proposed MDND framework, which introduces a paradigm shift by incorporating a non-differentiable iterative refinement oracle. This oracle generates a high-quality target map, while gradients for training the feature extractor propagate exclusively through the parallel, differentiable branch. This design decouples refinement from learning and avoids potential gradient conflicts.这张图概括 MDND 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 1 · : Qualitative comparison on challenging shapes**Figure 1: Qualitative comparison on challenging shapes. We visualize correspondences via texture transfer, comparing our method against state-of-the-art approaches like ULRSSM (Cao, Roetzer, and Bernard 2023) and HybridFMaps (Bastian et al. 2024). Our method produces noticeably more accurate and coherent maps in challenging scenarios involving non-isometric deformations (top rows) and significant topological noise (bottom row).这张图/表用于判断 MDND 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：用非可微 refinement branch 生成高质量内部目标，再监督可微 branch，无监督学习形状 correspondence。

方法拆解

用非可微 refinement branch 生成高质量内部目标，再监督可微 branch，无监督学习形状 correspondence

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。