Research map / Detection and segmentation / Open-set and few-shot detection

Multi-Sensor Object Anomaly Detection: Unifying Appearance, Geometry, and Internal Properties

CVPR 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Industrial anomaly detection · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MulSen-AD, a high-resolution RGB/infrared/point-cloud dataset of 15 industrial objects, plus MulSen-TripleAD, a decision-level fusion baseline of PatchCore-style memory banks for unsupervised object anomaly detection.

Problem

Single-sensor object anomaly detection cannot capture the full range of industrial anomalies (appearance, geometry, internal defects), and no multi-sensor benchmark existed for it.

Approach

Combine per-sensor PatchCore-style memory banks (DINO features for RGB and infrared, Point-MAE for point clouds) with a learnable one-class SVM decision gating unit (from M3DM) to fuse anomaly scores at the decision level, trained only on normal samples.

Key results

• MulSen-TripleAD (RGB+IR+PC) achieves 96.1% mean object-level AUROC on MulSen-AD, versus 91.1% for RGB alone, 90.9% for IR alone, and 66.8% for point cloud alone.
• Dual-sensor RGB+IR reaches 94.8% AUROC, and point cloud+RGB and point cloud+IR reach 91.9% AUROC.
• Pixel-level localization: RGB achieves mean pixel-AUROC 0.982, IR 0.970, point cloud 0.632.

Measured effects reported in the paper

Surprising findings

• Point cloud alone performs poorly (66.8% AUROC), and its pixel-level localization is weak (pixel-F1-max 0.111).
• Infrared-based and RGB-based single-sensor baselines are competitive with the fused model on several categories, with per-category ranking varying widely across methods.

Limitations

• Decision-level fusion may miss cross-modal interactions that feature-level fusion could capture.
• No X-ray or other deeper-sensing modality is included.
• Real-time performance and scalability for resource-constrained settings are not addressed.

Experimental setup

Datasets: MulSen-AD

Benchmarks: MulSen-AD

Metrics: object-level AUROC pixel-AUROC pixel-F1-max pixel-AUPR

Baselines: CFA CFLOW-AD DeSTSeg DRAEM InvAD PatchCore RD++ SimpleNet BTF M3DM Reg3D-AD

Builds on: PatchCore M3DM DINO Point-MAE

Closest papers by meaning

• Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.86)
• Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.86)
• Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics: Physics-Grounded Anomaly Detection (CVPR 2025 · similarity 0.86)
• MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images (ICLR 2024 · similarity 0.85)
• 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.85)
• AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.85)