DualBP-AD: A Dual-Path Back-Projection Framework for Universal 3D Anomaly Detection

3D anomaly detection serves as a crucial task in industrial applications, dedicated to identifying the anomalies and localizing the anomalous regions. However, existing methods are typically tailored to specific task configurations and particular modality inputs, thereby lacking a universal paradigm. To overcome these bottlenecks, we propose DualBP-AD, a universal dual-path back-projection framework that conceptualizes anomaly detection through a cross-modal rendering paradigm. By projecting unstructured 3D data into 2D images, our framework leverages the robust generalization capability of 2D foundational models , ensuring these extracted features maintain strong semantic power when back-projected onto 3D spatial points. To resolve contextual confusion across varying viewpoints and categories, we design an instance-adaptive spatial-semantic transition layer. Furthermore, we introduce an orthogonal dual-path back-projection mechanism to accurately map and fuse multi-view anomaly back into the native 3D topology. Extensive experiments demonstrate that DualBP-AD overcomes task-specific and data-specific limitations, yielding highly accurate detection and precise localization across single-class, multi-class, few-shot, and zero-shot settings on four benchmarks of diverse modalities. For example, it achieves 95.6% O-AUROC and 97.3% P-AUROC on Anomaly-ShapeNet, outperforming the previous SOTA.