Source-linked AI summary
GeoMAD: Geometry-Aware Multi-View Anomaly Detection via Deformable Fusion and Distributional Alignment
Shang-Fu Chen, Jhih-Ciang Wu, Kuan-Chuan Peng, Wen-Huang Cheng, Kai-Lung Hua
TL;DR
MvAD needs scalable fusion that captures cross-view geometry without the costs and assumptions of explicit 3D construction, while also addressing distributional inconsistency. GeoMAD combines CDFM for adaptive 2D correspondence with DVA for cross-view distribution alignment, and improves unified detection and localization on Real-IAD and MANTA-Tiny. Its training-time DVA regularizes the fusion bottleneck, while inference retains per-view teacher–student anomaly evidence.
Problem
MvAD must fuse complementary views geometrically while scaling to multiple classes, but existing approaches have costly 3D assumptions or discrete patch-level correspondence.
Method
GeoMAD uses CDFM for calibration-free, content-adaptive cross-view sampling and DVA to align each bottleneck distribution with a per-instance view-centric target.
Results
GeoMAD improves unified multi-view anomaly detection and localization on Real-IAD and MANTA-Tiny over existing methods.
Takeaways & Limitations
CDFM and DVA combine local deformable correspondence with global view-level consistency without explicit 3D reconstruction.
Takeaways & Limitations
DVA constrains the cross-fusion bottleneck only during training and does not modify inference anomaly maps derived from per-view teacher–student divergence.
Abstract
from arXiv · showhide
Multi-view anomaly detection (MvAD) detects defects by exploiting complementary observations from multiple camera viewpoints. The central challenge is to fuse views with sufficient geometric awareness while remaining scalable to multi-class industrial settings. Existing methods typically fall into two extremes: voxel-based fusion provides explicit geometric alignment but requires costly 3D construction and class-specific assumptions, whereas lightweight patch-based fusion is efficient but relies on discrete candidate matching and lacks continuous cross-view correspondence. In this paper, we propose GeoMAD, a unified multi-view, multi-class AD framework that addresses both geometric correspondence deficiency and distributional inconsistency. Our \textit{Cross-view Deformable Fusion Module} (CDFM) learns content-adaptive, view-pair-specific sampling offsets directly on 2D feature maps and arranges them across a multi-scale window pyramid with image-global reference sampling, enabling hierarchical cross-view correspondence without camera calibration, voxel construction, or class-specific 3D supervision. We further introduce \textit{Distributional View Alignment} (DVA), a self-supervised cross-view regularization loss that aligns each view's bottleneck distribution against a per-instance view-centric target, enforcing global consistency without pixel-level correspondence. Together, CDFM and DVA bridge local geometric correspondence and global distributional consistency, providing geometry-aware and distribution-consistent fusion while preserving the efficiency of 2D feature-space learning. Extensive experiments on Real-IAD and MANTA-Tiny show that GeoMAD achieves strong detection and localization performance in unified MvAD.
1 Introduction
GeoMAD targets the geometric and distributional challenges of scalable multi-view, multi-class anomaly detection. It combines deformable 2D cross-view fusion with distributional alignment and reports improved unified detection and localization.
- Multi-view anomaly detection uses complementary synchronized views because defects can be occluded, poorly illuminated, or geometrically ambiguous from one viewpoint.
- Existing methods trade off scalability and correspondence quality: voxel fusion requires costly 3D construction and class-specific assumptions, while patch fusion uses discrete candidates.
- GeoMAD jointly addresses adaptive cross-view correspondence and view-level distributional consistency without explicit 3D reconstruction.
- CDFM learns calibration-free, content-adaptive, view-pair-specific sampling offsets with image-global reach across a multi-scale window pyramid.
- DVA aligns each view’s bottleneck distribution with a per-instance view-centric target without anomaly labels, pixel-level correspondence, or explicit 3D projection.
- Experiments on Real-IAD and MANTA-Tiny show improved unified multi-view anomaly detection and localization over existing methods.
2 Related Work
Related work spans class-specific and unified anomaly detection, with multi-view approaches differing in computational design and geometric alignment. GeoMAD combines learned 2D correspondence and distributional alignment while avoiding explicit 3D reconstruction.
- Class-specific anomaly detection trains separate models per object category, achieving strong performance but increasing storage and deployment costs.
- Unified anomaly detection uses one model across categories, improving scalability while introducing potential limitations described in prior work.
- GeoMAD combines learned 2D cross-view correspondence through CDFM with distributional view alignment through DVA, avoiding explicit 3D reconstruction for multi-class anomaly detection.
3 Our proposed Method – GeoMAD
GeoMAD processes synchronized multi-view inputs with a reverse-distillation pipeline that fuses views at multiple feature scales and combines reconstruction with distributional consistency. Its CDFM provides continuous, view-pair-specific cross-view sampling, while DVA aligns bottleneck distributions across views without explicit 3D supervision.
- 3.1 Overview: GeoMAD organizes anomaly-free training instances as synchronized sets of V views and reshapes the batch accordingly for multi-view processing.The teacher extracts multi-scale features, while the student reconstructs those scales after cross-view bottleneck fusion.
- 3.1 Overview: The pipeline combines multi-scale reverse-distillation loss with DVA, and inference aggregates dense pixel-level, image-level, and sample-level anomaly scores.Multi-scale cosine distances support pixel-level maps; spatial pooling produces view scores, followed by averaging across viewpoints for sample scores.
- 3.2 Cross-view Deformable Fusion Module: CDFM replaces fixed cross-view locations with content-adaptive, view-pair-specific offsets predicted from query features and applied to reference features.Separate offset fields model different camera-pair displacements while sharing the offset predictor.
- 3.2 Cross-view Deformable Fusion Module: Bilinear sampling retrieves reference keys and values at continuous deformed positions, reducing boundary fragmentation from rigid patch partitioning.Windowed queries retain global reference-view key/value reach through deformable sampling.
- 3.2 Cross-view Deformable Fusion Module: CDFM applies cross-view fusion across a multi-scale window pyramid, then aggregates responses from all reference views into fused bottleneck representations.Uniform aggregation provides a view-symmetric prior while learned offsets determine where reference views contribute.
- 3.3 Distributional View Alignment: DVA aligns each view’s bottleneck distribution to a per-instance target formed from the views themselves, complementing CDFM’s local correspondence with global consistency.The regularization is applied during training and does not modify inference anomaly maps derived from teacher–student divergence.
4 Experiments
Experiments evaluate GeoMAD on Real-IAD and MANTA-Tiny using unified multi-class settings, quantitative comparisons, ablations, robustness tests, and efficiency analysis. Results show strong detection and localization, with deformable fusion and distributional alignment contributing to performance and qualitative coherence.
- Qualitative Comparison: Qualitative comparisons show more spatially coherent localization and fewer false activations for GeoMAD across five views of each object instance.The reported behavior contrasts with MVAD’s patch-boundary defect fragmentation and background over-activation.
- Comparison with the SOTA methods: GeoMAD ranks first on 8 of 10 Real-IAD metrics and second on the remaining two, improving S-AUROC by 1.6, P-AP by 5.3, and P-F1 by 4.0 points over MVAD.DeSTSeg leads on P-AP and P-F1 but uses synthetic anomalies and pseudo-ground-truth masks.
- Ablation Study: Adding CDFM produces substantial gains on both datasets, while adding DVA further improves performance across most metrics on both datasets.The ablation identifies deformable cross-view fusion as a strong inductive bias for learning correspondence without explicit 3D supervision.
- Robustness to Missing Views: Inference with progressively fewer views, from 5 to 2, is evaluated at sample-, image-, and pixel-level metrics to assess missing-view robustness.The accompanying discussion states that image- and pixel-level performance remain highly stable as views are removed.
- Computational Efficiency: GeoMAD’s non-voxel design is reported as efficient relative to voxel-related modules and comparable architectures.Computational efficiency is assessed using parameters, FLOPs, and training memory.
5 Conclusion
GeoMAD combines calibration-free deformable cross-view correspondence with distributional view alignment in a unified 2D architecture. Experiments on Real-IAD and MANTA-Tiny show improved unified multi-view anomaly detection and localization, while future work could address dynamic view configurations.
- 5 Conclusion: GeoMAD addresses geometric correspondence deficiency and distributional inconsistency using CDFM and DVA without explicit 3D reconstruction.CDFM learns content-adaptive, view-pair-specific offsets; DVA aligns bottleneck distributions to a stop-gradient view-centric target.
- 5 Conclusion: Experiments on Real-IAD and MANTA-Tiny show improved unified multi-view anomaly detection and localization over existing methods.The conclusion supports jointly addressing geometric and distributional gaps in cross-view learning.
- 5 Conclusion: Future work could extend GeoMAD to handle dynamic view configurations.