Research map / Label-efficient and robust learning / OOD and anomaly detection
TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection
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 TailedCore, a memory-based anomaly detector that adds tail-class patch features selected by a symmetric-similarity class size predictor (TailSampler) to a noise-filtered SoftPatch-style memory, targeting unsupervised long-tail noisy anomaly detection.
Problem
Unsupervised anomaly detection on multi-class product training sets that are both contaminated with defective samples and have long-tailed, unlabeled class distributions, where existing methods trade off tail-class performance against noise robustness.
Approach
Estimate each training sample's class size from a reflective symmetry between inter-class and intra-class embedding similarity (adaptive half-angle neighborhoods plus neighborhood-of-neighborhood mode voting, with an elbow rule for the tail-size cutoff), then sample tail-class patches exclusively and union them with a noise-filtered coreset.
Key results
- On MVTecAD with step (K=1) tail distribution, TailedCore reaches 95.72% image-level AUROC overall versus 90.37% for WinCLIP and 80.01% for PatchCore.
- On VisA with Pareto tail distribution, TailedCore reaches 90.85% image-level AUROC overall versus 85.59% for PatchCore and 86.56% for SoftPatch.
- TailSampler achieves a combined missing-tail plus included-anomaly ratio of 5.10% versus 18.15% for Isolation Forest in the few-shot sampling comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TailedCore instead of the strongest baseline | unsupervised long-tail noisy anomaly detection · MVTecAD, Pareto, step (K=4), and step (K=1); all classes | image-level AUROC (%): 91.48 -> 96.12; 91.48 -> 95.71; 91.48 -> 94.43 | Helped |
| Use TailedCore instead of the strongest baseline | unsupervised long-tail noisy anomaly detection · VisA, Pareto, step (K=4), and step (K=1); all classes | image-level AUROC (%): 86.56 -> 90.85; 82.05 -> 89.64; 82.05 -> 87.61 | Helped |
| Use TailedCore instead of the strongest baseline | unsupervised long-tail noisy anomaly detection · MVTecAD, Pareto, step (K=4), and step (K=1); all classes | pixel-level AUROC (%): 93.41 -> 95.29; 91.53 -> 94.74; 91.08 -> 93.99 | Helped |
| Use TailedCore instead of the strongest baseline | unsupervised long-tail noisy anomaly detection · VisA, Pareto, step (K=4), and step (K=1); all classes | pixel-level AUROC (%): 96.71 -> 97.48; 95.74 -> 96.89; 95.51 -> 96.65 | Helped |
| Add few-shot augmentation (FSA) to TailedCore's memory (ablation) | TailedCore · MVTecAD, all tail types; all classes | image-level AUROC (%): 88.13 -> 96.13; 75.05 -> 94.44; 78.94 -> 95.72 | Helped |
| Add few-shot augmentation (FSA) to TailedCore's memory (ablation) | TailedCore · MVTecAD, all tail types; all classes | pixel-level AUROC (%): 93.54 -> 95.30; 80.14 -> 94.00; 86.23 -> 94.75 | Helped |
| Set TailSampler neighborhood percentile p to 100 rather than its default 85 (ablation) | TailSampler · MVTecAD and VisA, all tail distributions and seeds | AUROC (%): 99.78 ± 0.30 -> 99.76 ± 0.29 | No clear effect |
| Set TailSampler neighborhood percentile p to 50 rather than its default 85 (ablation) | TailSampler · MVTecAD and VisA, all tail distributions and seeds | AUROC (%): 99.78 ± 0.30 -> 97.36 ± 5.65 | Hurt |
| Use the nearest embedding's neighborhood size instead of majority voting (ablation) | TailSampler class-size predictor · MVTecAD and VisA, all tail distributions and seeds | AUROC (%): 99.78 ± 0.30 -> 99.73 ± 0.34 | No clear effect |
| Average neighborhood sizes instead of majority voting (ablation) | TailSampler class-size predictor · MVTecAD and VisA, all tail distributions and seeds | AUROC (%): 99.78 ± 0.30 -> 99.35 ± 2.59 | No clear effect |
Surprising findings
- Classical clustering and outlier detectors (e.g., LOF, Isolation Forest, DBSCAN) largely fail at few-shot sampling, and the noise-discrimination step in SoftPatch removes most few-shot class patches.
- Foundation-model zero-shot baselines (WinCLIP, AnomalyCLIP) are competitive with memory-based methods on several tail-heavy settings, where PatchCore and SoftPatch degrade.
Limitations
- Class size prediction fails when the reflective-symmetry assumption breaks down, e.g., with poor or label-misaligned encoder embeddings.
- Defect geometry can resemble few-shot class instances in embedding space, leading to incorrectly sampled tail patches and some head-class performance drop.
- Performance depends on the discriminativeness of the encoder embedding.
Experimental setup
Datasets: MVTecAD VisA
Benchmarks: MVTecAD VisA
Metrics: image-level AUROC pixel-level AUROC combined mis-sampling ratio class size prediction error
Baselines: PaDiM HVQ WinCLIP AnomalyCLIP PatchCore SoftPatch
Builds on: PatchCore SoftPatch WideResNet-50
Closest papers by meaning
- Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.88)
- Weakly Supervised Anomaly Detection via Dual-Tailed Kernel (ICML 2025 · similarity 0.87)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.86)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.85)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.84)
- We Need to Rethink Benchmarking in Anomaly Detection (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.