Research map / Label-efficient and robust learning / OOD and anomaly detection
Salvaging the Overlooked: Leveraging Class-Aware Contrastive Learning for Multi-Class Anomaly Detection
Lei Fan, Junjie Huang, Donglin Di, Anyang Su, Tianyou Song, Maurice Pagnucco, Yang Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add local CL (ablation) | RD (multi-class baseline) · MVTec+VISA+BTAD (all-in-all) | I-AUROC/P-AUROC/PRO: +5.5/+2.1/+2.4 | Improved |
| add global CL (ablation) | RD (multi-class baseline) · MVTec+VISA+BTAD (all-in-all) | I-AUROC/P-AUROC/PRO: +4.2/+1.6/+2.0 | Improved |
| combine local and global CL (full CCL, raw class labels) (ablation) | RD (multi-class baseline) · MVTec+VISA+BTAD (all-in-all) | I-AUROC/P-AUROC/PRO: +5.7/+2.2/+2.8 | Improved |
| use K-means pseudo-class labels (K_C=30) instead of raw class labels (ablation) | CCL · MVTec+VISA+BTAD (all-in-all) | I-AUROC/P-AUROC/PRO: +5.7/+2.2/+2.8 | No clear effect |
| use pseudo-classes with K_C=60 instead of 30 (ablation) | CCL · MVTec+VISA+BTAD (all-in-all) | I-AUROC/P-AUROC/PRO: +5.0/+2.0/+2.2 | Hurt |
| add more feature scales to local CL (f3 -> {f2,f3} -> {f1,f2,f3}) (ablation) | RD + local CL · MVTec+VISA+BTAD (all-in-all) | I-AUROC: +2.9 -> +3.7 -> +5.1 | Improved |
| use 2-layer MLP projector instead of 4 conv blocks (ablation) | RD + local CL · MVTec+VISA+BTAD (all-in-all) | I-AUROC/P-AUROC/PRO: -15.3 / -8.9 / -23.8 | Hurt |
| use 4 conv blocks projector instead of 2 (ablation) | RD + local CL · MVTec+VISA+BTAD (all-in-all) | I-AUROC: +4.7 -> +5.1 | Improved |
| shrink local window k from max(h,w) to 1 (ablation) | RD + local CL · MVTec+VISA+BTAD (all-in-all) | I-AUROC: +5.1 -> +5.5 | Improved |
| replace triplet loss with L_gcl as global loss (ablation) | RD + global CL · MVTec+VISA+BTAD (all-in-all) | I-AUROC: 91.0 -> 95.2 | Improved |
| use all same-class samples as positives (L_gcl) instead of InfoNCE (ablation) | RD + global CL · MVTec+VISA+BTAD (all-in-all) | I-AUROC: 93.9 -> 95.2 | Improved |
| vary lambda1:lambda2 ratio (0.2 to 2) in CCL (ablation) | CCL · MVTec+VISA+BTAD (all-in-all) | I-AUROC: 96.7 (1:1) vs 96.5 (2) | No clear effect |
Closest papers by meaning
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.87)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.87)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.87)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.87)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.86)
- CNC: Cross-modal Normality Constraint for Unsupervised Multi-class Anomaly Detection (AAAI 2025 · similarity 0.86)
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.
