MATCH: Flow Matching for Multi-View Anomaly Detection

A new multi-view anomaly detection method called MATCH achieves state-of-the-art results on industrial inspection benchmarks while running on consumer-level hardware, according to a paper submitted to arXiv in June 2026 [1][2]. The method, detailed in a preprint from the Computer Vision and Pattern Recognition section of arXiv, is the first to apply Flow Matching (FM) to the problem of spotting defects across multiple camera angles of manufactured objects [1][2]. Flow Matching belongs to a broader family of generative models that includes diffusion models, which learn to reverse a process of adding noise to data [6]. MATCH uses the ordinary differential equation (ODE) formulation of FM to estimate likelihoods and derive an anomaly score at the object, image, and pixel levels [1][2]. The authors evaluated MATCH on the established Real-IAD dataset and provided what they describe as the first comprehensive multi-view anomaly detection benchmark on the MANTA-Tiny dataset [1][2]. High-quality labeled datasets are typically difficult and expensive to produce for supervised machine learning, making benchmark contributions a significant part of research progress [7]. The paper reports that MATCH achieves state-of-the-art performance in both anomaly detection and segmentation tasks [1][2]. A key architectural choice was omitting the computationally expensive divergence term normally required for likelihood estimation in flow-based models [1][2]. This decision, the authors write, ensures MATCH is usable in real-time production scenarios [1][2]. The model's ability to run on consumer-level hardware further lowers the barrier for deployment on factory floors [1][2]. Anomaly detection itself is a well-established concept beyond manufacturing. In network security, for example, anomaly-based intrusion detection systems monitor traffic for deviations from a model of normal behavior, often relying on machine learning [3]. The MATCH approach applies a conceptually similar principle—learning the distribution of normal multi-view features and flagging deviations—to the visual inspection domain [1][2]. The preprint was posted on arXiv, the open-access e-print repository that hosts scientific papers across physics, computer science, and other fields [11]. As of late 2024, arXiv was receiving about 24,000 submissions per month [11]. The paper includes ablation studies to validate the specific methodological choices behind MATCH's architecture [1][2].