CELAD: Compositional Evaluation for Logical Anomaly Detection

ICLR 2026 Conference Submission13784 Authors

Keywords: Computer Vision, Anomaly Detection, Logical Anomaly Detection

TL;DR: We introduce CELAD, a benchmark for logical anomaly detection emphasizing compositionality, and ROMAD, a new method that outperforms prior approaches on CELAD by 15% while maintaining strong performance on existing datasets.

Abstract: Anomaly detection (AD) has attracted significant research interest and now achieves near-perfect performance on most existing benchmarks. However, the majority of prior work has focused on detecting structural anomalies, where anomalies manifest as localized defective regions. Recently, logical anomaly detection (LAD) has emerged, extending AD to cases where violations occur at the level of compositional or relational rules rather than individual regions—a setting particularly relevant for industrial inspection. Despite its importance, LAD research remains hindered by limited dedicated datasets, raising concerns about the generalization ability of current methods. We address this gap with two contributions. First, we introduce CELAD, a new benchmark designed to test compositional understanding in LAD. CELAD features greater variation in both normal and anomalous samples, along with more intricate anomaly types, resulting in a substantial performance drop in state-of-the-art methods. Second, we propose ROMAD, a simple yet effective framework that leverages DETR, an object detector with strong relational modeling, to construct rich object embeddings. ROMAD computes anomaly scores via a training-free matching pipeline and requires only a small number of annotated samples. Extensive experiments show that ROMAD achieves a new state of the art on CELAD, outperforming the next-best method by nearly 15% while maintaining competitive performance on existing datasets. In few-shot regimes, ROMAD further delivers the strongest average results across both CELAD and prior benchmarks, demonstrating its ability to generalize beyond memorization and capture the underlying logical rules. Code and data are available at: https://github.com/neutral-coder-737/Home-Page.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 13784