PCAD‐Flow: Point Clouds Anomaly Detection via Normalizing Flow

Unsupervised 3D point clouds anomaly detection aims to identify and localize anomalous geometric regions only using normal training samples. Existing methods predominantly rely on feature matching and geometric reconstruction, or implicit shape modeling, where anomalies are indirectly identified through task‐specific deviation measures, such as feature discrepancies, reconstruction errors, or geometric deviations. Different from previous works, in this paper, we propose a new flow‐based framework (PCAD‐Flow) for 3D point clouds anomaly detection, which directly evaluates the anomaly scores from the point clouds by utilizing point‐wise features based on exact likelihood estimation. Specifically, PCAD‐Flow employs conditional normalizing flows to learn invertible mappings from point‐wise feature representations to a tractable latent distribution, allowing anomaly scores to be computed directly from negative log‐likelihoods under the learned normal distribution. To further capture local geometric structure, we introduce a Progressive Local Conditioning (PLC) module that progressively aggregates neighborhood information and injects multi‐stage contextual cues into the flow transformation. This context‐aware density modeling enables the framework to effectively distinguish anomalous geometric patterns from normal local variations. Extensive experiments on Real3D‐AD and Anomaly‐ShapeNet demonstrate that PCAD‐Flow achieves state‐of‐the‐art performance on both object‐level anomaly detection and point‐level anomaly localization. Particularly, the proposed method obtains substantial improvements in localization accuracy than previous methods on the real‐world point clouds benchmarks.