CoGeoAD: Hierarchical Color-Geometric Fusion with Multi-View Attention for Zero-Shot 3D Anomaly Detection

Industrial anomaly detection often faces a scarcity of labeled data, making zero-shot approaches essential. While 2D-RGB methods excel at texture detection and 3D point cloud methods excel at structural analysis, existing frameworks struggle to effectively integrate these complementary modalities. CoGeoAD aims to bridge this gap by creating a unified, CLIP-based framework that performs hierarchical fusion of color and geometric features without requiring target-domain training.

The authors propose a three-stage pipeline:

1. Multi-View Rendering: The system projects 3D point clouds into paired 2D color and geometric images. It uses a pinhole camera model and visibility masks to establish precise point-to-pixel correspondences, effectively filtering out occlusion noise.
2. Data-Driven Multi-View Attention (MVA): A lightweight CNN-based encoder computes importance scores for each viewpoint. This allows the model to dynamically prioritize informative views and suppress noise, rather than relying on fixed weighting schemes.
3. Multi-Stage Color-Geometric Fusion (MS-CGF): The framework extracts multi-level features from the frozen CLIP encoder. It uses a dynamic layer selector to aggregate features across different semantic depths and computes anomaly scores for color and geometry in parallel. The final anomaly map is generated via an element-wise maximum operation, ensuring that strong evidence from either modality is preserved.

CoGeoAD demonstrates that zero-shot 3D anomaly detection can be significantly improved by treating color and geometry as distinct but synergistic inputs. By avoiding internal modifications to the pre-trained CLIP encoder, the model preserves its foundational generalizability while achieving superior sensitivity to both subtle surface stains and complex structural cracks. This approach provides a robust, training-free solution for industrial inspection scenarios where data imbalance is a major bottleneck.