GPFlow：面向无监督多模态异常检测的高斯原型概率流
GPFlow: Gaussian Prototype Probability Flow for Unsupervised Multi-Modal Anomaly Detection
用可学习高斯原型构造协方差感知的信息瓶颈（后验均值路径路由），替代离散原型记忆库，可直接迁移到少样本工业质检、多模态（RGB+点云/深度）异常检测的特征重建环节。
原文依据 an analytical Posterior-Mean Path (PMP) router, which reconstructs features through the posterior mean of a noise-smoothed Gaussian mixture
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论本文研究在具有挑战性的少样本条件下的无监督多模态异常检测，即每个类别仅有少量正常训练样本可用。为避免对异常进行平凡的重建，近期方法通常依赖离散原型来建立信息瓶颈。然而，此类离散原型无法刻画正常特征的连续且各向异性的变化。为此，我们提出 GPFlow，一个受概率流启发的框架，使用可学习的高斯原型对正常性进行建模。GPFlow 的关键组件是解析式的后验均值路径（Posterior-Mean Path, PMP）路由器，它通过噪声平滑高斯混合模型的后验均值来重建特征。这产生了作为协方差感知信息瓶颈的各向异性收缩：PMP 有选择地保留与高斯原型协方差结构对齐的正常变化，同时严格抑制与原型不一致的偏差。为利用跨模态的互补知识，GPFlow 进一步结合模态内与跨模态重建，并施加轻量的实例感知先验校准，以缓解稀疏训练数据与多样测试样本之间的分布失配。在 MVTec-3D-AD 与 Eyecandies 上的实验表明，GPFlow 仅需少量正常训练样本即可取得显著的性能提升，同时保持计算高效。
英文原文摘要 1442 字符
In this paper, we study unsupervised multi-modal anomaly detection under challenging few-shot conditions, where only a few normal training samples are available for each class. To prevent the trivial reconstruction of anomalies, recent methods often rely on discrete prototypes to establish an information bottleneck. However, such discrete prototypes fail to capture continuous and anisotropic variations of normal features. We therefore propose GPFlow, a probability-flow-inspired framework that models normality using learnable Gaussian prototypes. The key component of GPFlow is an analytical Posterior-Mean Path (PMP) router, which reconstructs features through the posterior mean of a noise-smoothed Gaussian mixture. This yields anisotropic shrinkage as a covariance-aware information bottleneck: PMP selectively preserves normal variations aligned with the covariance structure of Gaussian prototypes while strictly suppressing deviations inconsistent with the prototypes. To exploit complementary knowledge across modalities, GPFlow further combines intra-modal and cross-modal reconstruction, and applies a lightweight instance-aware prior calibration to alleviate the distribution mismatch between sparse training data and diverse test samples. Experiments on MVTec-3D-AD and Eyecandies show that GPFlow achieves significant performance improvement with only a few normal training samples while remaining computationally efficient.