MATCH: Flow Matching for Multi-View Anomaly Detection

In industrial manufacturing, detecting anomalies across multiple viewpoints of complex objects is critical for quality control. Existing methods often struggle with high-dimensional image data or the computational overhead required for accurate density estimation. This paper introduces MATCH, a novel approach designed to perform multi-view anomaly detection efficiently while maintaining high precision in both image-level detection and pixel-level segmentation.

MATCH leverages Flow Matching (FM), a generative modeling technique that learns to transform data distributions into a simple Gaussian distribution via an ordinary differential equation (ODE). Unlike traditional Normalizing Flows that rely on restrictive coupling architectures, Flow Matching allows for more flexible network designs. The authors adapt a Reverse Distillation architecture, using a frozen WideResNet as a feature encoder. To ensure the model is suitable for real-time production environments, the authors propose a key simplification: they omit the costly divergence term typically required for exact likelihood estimation. Instead, they use the squared norm of the latent representation in Gaussian space as a proxy for the anomaly score, which is computationally efficient and maintains high performance.

MATCH achieves state-of-the-art results on the Real-IAD and MANTA-Tiny datasets for both anomaly detection and segmentation. By removing the divergence term, the model significantly reduces computational requirements, enabling it to run at approximately 18.77 FPS on consumer-grade hardware (RTX 3090). The authors demonstrate that this simplification does not harm the model's ability to distinguish between normal and anomalous samples, making it a practical solution for industrial deployment.