SuperAD: A Training-free Anomaly Classification and Segmentation Method for CVPR 2025 VAND 3.0 Workshop Challenge Track 1: Adapt & Detect

TL;DR

This work tackles robust, training-free anomaly detection and segmentation for real-world industrial data under complex lighting and occlusion. It builds a compact memory bank of 16 normal references per category using greedy coreset selection on DINOv2 CLS features and performs per-pixel nearest-neighbor matching across four feature levels to generate anomaly maps, which are fused into a final segmentation. The approach demonstrates competitive performance on the MVTec AD 2 dataset and superior robustness under varying illumination, without any model fine-tuning, highlighting the strong generalization power of large self-supervised representations. Practical impact lies in providing a lightweight, training-free solution suitable for deployment in diverse industrial environments with limited labeled anomalies.

Abstract

In this technical report, we present our solution to the CVPR 2025 Visual Anomaly and Novelty Detection (VAND) 3.0 Workshop Challenge Track 1: Adapt & Detect: Robust Anomaly Detection in Real-World Applications. In real-world industrial anomaly detection, it is crucial to accurately identify anomalies with physical complexity, such as transparent or reflective surfaces, occlusions, and low-contrast contaminations. The recently proposed MVTec AD 2 dataset significantly narrows the gap between publicly available benchmarks and anomalies found in real-world industrial environments. To address the challenges posed by this dataset--such as complex and varying lighting conditions and real anomalies with large scale differences--we propose a fully training-free anomaly detection and segmentation method based on feature extraction using the DINOv2 model named SuperAD. Our method carefully selects a small number of normal reference images and constructs a memory bank by leveraging the strong representational power of DINOv2. Anomalies are then segmented by performing nearest neighbor matching between test image features and the memory bank. Our method achieves competitive results on both test sets of the MVTec AD 2 dataset.

Figures (2)

• Figure 1: The overall architecture of our proposed method.
• Figure 2: Examples of typical failure cases: (a) false positives on air bubbles; (b) missed detection of blurred objects; (c) false positives due to specular highlights; (d) missed detections of missing-type anomalies.