MVFM-3DAD: Multi-view Flow Matching for 3D Anomaly Detection via Density Proxy Estimation

MVFM-3DAD reframes 3D anomaly detection as density proxy estimation instead of retrieval or reconstruction. That matters because memory-bank systems can miss normal cases they never stored, while reconstruction models can learn to reproduce anomalies too well. The new pipeline aims to judge whether a test sample belongs to the normal distribution directly, which is a cleaner fit for inspection-style 3D pipelines.

The core of the system is a Bidirectional Geometric Projector. Its forward pass turns irregular point clouds into structured multi-view representations, then a Flow-guided Density Proxy Estimator assigns a reference density to each view feature. The backward pass maps those density estimates back to the 3D points, so anomalous features are identified by how far they drift from terminal normality.

For developers, the practical appeal is simplicity in scoring and fewer brittle assumptions. The method does not require input reconstruction or explicit Jacobian evaluation, which can make flow-based approaches expensive or awkward to deploy. In the reported experiments, MVFM-3DAD outperformed the strongest competing methods on Real3D-AD and MVTec3D-AD, two common benchmarks for 3D anomaly detection.

This is still research territory, but it points toward a more robust way to handle inspection of scanned assets, captured geometry, or other point-cloud-heavy workflows. If the approach holds up outside benchmark data, it could be useful anywhere teams need to flag outliers in 3D content without overfitting to a narrow library of known-good shapes.