Visual Anomaly Detection under Complex View-Illumination Interplay: A Large-Scale Benchmark

TL;DR

The paper targets the real-world gap in visual anomaly detection (VAD) by focusing on the complex interplay between viewpoint and illumination. It introduces M2AD, a large-scale benchmark with synchronized multi-view and multi-illumination data, consisting of 120 configurations per specimen across 999 specimens and 10 categories, totaling 119,880 high-resolution images. Two evaluation protocols, M2AD-Synergy and M2AD-Invariant, enable rigorous assessment of multimodal fusion and single-image robustness under realistic imaging conditions. Benchmarking shows that state-of-the-art unsupervised VAD methods experience substantial performance drops on M2AD, underscoring the need for fusion-aware, high-resolution, and physics-informed approaches. The dataset and evaluation framework provide a practical foundation for developing robust VAD systems suited to industrial inspection and other real-world applications.

Abstract

The practical deployment of Visual Anomaly Detection (VAD) systems is hindered by their sensitivity to real-world imaging variations, particularly the complex interplay between viewpoint and illumination which drastically alters defect visibility. Current benchmarks largely overlook this critical challenge. We introduce Multi-View Multi-Illumination Anomaly Detection (M2AD), a new large-scale benchmark comprising 119,880 high-resolution images designed explicitly to probe VAD robustness under such interacting conditions. By systematically capturing 999 specimens across 10 categories using 12 synchronized views and 10 illumination settings (120 configurations total), M2AD enables rigorous evaluation. We establish two evaluation protocols: M2AD-Synergy tests the ability to fuse information across diverse configurations, and M2AD-Invariant measures single-image robustness against realistic view-illumination effects. Our extensive benchmarking shows that state-of-the-art VAD methods struggle significantly on M2AD, demonstrating the profound challenge posed by view-illumination interplay. This benchmark serves as an essential tool for developing and validating VAD methods capable of overcoming real-world complexities. Our full dataset and test suite will be released at https://hustcyq.github.io/M2AD to facilitate the field.

Figures (20)

• Figure 1: Motivation.(a) Anomaly detectability is governed by complex view-illumination interplay. Each image pair shows the original input (left) alongside its corresponding ground truth (right), with anomaly regions highlighted in white. (b) To address this challenge, our M$^2$AD introduces multi-view and multi-illumination acquisition protocols, enabling robust anomaly detection across diverse conditions. Zoom in for a clearer view. More samples in M$^2$AD are visualized in Appendix Sec. \ref{['supp:dataset_samples']}.
• Figure 2: Data collection pipeline of M$^2$AD. A three-step process is employed. (a) Object preparation and defect engineering. (b) Design and construction of a configurable imaging prototype capable of capturing multi-view, multi-illumination images programmatically. (c) Assessing detectability by evaluating the consistency between predictions and annotations for M$^2$AD-Invariant.
• Figure 3: Statistics of M$^2$AD. (a) Distribution of normal and abnormal object, view, and image counts across different categories. "Detectable" refers to the abnormal images retained in Sec. \ref{['sec:data_acquistion']}. (b) Percentage of image area occupied by anomaly regions. (c) Aspect ratio statistics of the minimum bounding rectangle of defects.
• Figure 4: Ablation study results. (a) O-AUROC under different imaging configuration numbers (24, 48, 72, 96, and 120). (b) I-AUROC and AUPRO under different combinations of illumination conditions (2, 4, 6, 8, and 10). We randomly select the configurations and illumination conditions three times and report (mean $\pm$ std).
• Figure 5: Visualization of anomaly detection results. (a) Input image, (b) ground truth (anomalies highlighted in white), (c) predicted anomaly maps by the best-performing model Dinomaly dinomaly. Despite its robustness, the visualization demonstrates Dinomaly’s limitations in capturing anomalies across diverse scenarios. Zoom in for a clearer view. See Appendix Sec. \ref{['supp:anomaly_visualization']} for more visualizations.