Research map / Label-efficient and robust learning / OOD and anomaly detection

CostFilter-AD: Enhancing Anomaly Detection through Matching Cost Filtering

ICML 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class anomaly detection · 0 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.

Proposes CostFilter-AD, a plug-in that builds a global patch-matching cost volume between an input and normal templates and refines it with a 3D U-Net guided by input features and an initial anomaly map, improving multi-class and single-class UAD on MVTec-AD and VisA.

Problem

Unsupervised anomaly detection relies on image- or feature-level matching between inputs and normal templates, and this matching is noisy (identical shortcut, misalignment), blurring normal/anomaly boundaries and hurting detection of subtle anomalies.

Approach

Treat anomaly scoring as cost volume filtering borrowed from stereo/flow estimation: build a global matching cost volume across multiple templates (including intermediate diffusion reconstructions), then filter it with a 3D U-Net whose residual channel-spatial attention is queried by input features and an initial anomaly map, plus a class-aware adaptor in the loss.

Key results

• Integrated with GLAD on MVTec-AD multi-class, image/pixel AUROC improves by 1.2%/0.9% (GLAD baseline 97.5/97.3 to 98.7/98.2 mean).
• On VisA, AnomalDF+Ours reaches mean 94.3 I-AUROC / 99.2 P-AUROC versus 90.5 / 97.5 for AnomalDF.
• With Dinomaly on MVTec-AD, CostFilter-AD achieves 99.7 I-AUROC and 98.4 P-AUROC versus 99.6 / 98.3 baseline, with only ~43.6M extra parameters and ~0.05 s extra inference per image for Dinomaly.

Measured effects reported in the paper

Surprising findings

• Directly applying stereo-matching-style local (pixel-level) matching drops performance to 87.8/89.0 on MVTec-AD, attributed to feature contamination from global multi-template matching.
• Anomaly volumes from reconstruction-based and embedding-based templates are not compatible across models; cross-type testing causes performance drops, requiring a hybrid model.

Limitations

• Relies on anomaly-relevant signals being present in the cost volume; subtle anomalies from low-resolution or weak features are still missed (failure cases).
• Gains over strong baselines like Dinomaly are small (e.g., 0.1 I-AUROC on VisA), and some per-class metrics regress.
• Cost volume compatibility is limited across template types, needing a hybrid model.

Experimental setup

Datasets: MVTec-AD VisA MPDD BTAD

Benchmarks: MVTec-AD VisA MPDD BTAD

Metrics: I-AUROC I-AP I-F1max P-AUROC P-AP P-F1max AUPRO

Baselines: UniAD GLAD HVQ-Trans AnomalDF Dinomaly PatchCore DiAD VPDM MambaAD JNLD OmniAL

Builds on: GLAD AnomalyDINO HVQ-Trans DINO stereo matching cost volume filtering

Closest papers by meaning

• One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.87)
• Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.87)
• AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.86)
• Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.86)
• UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.86)
• ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.85)