Research map / Label-efficient and robust learning / OOD and anomaly detection
UniNet: A Contrastive Learning-guided Unified Framework with Feature Selection for 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 UniNet, a unified student-teacher anomaly detection framework with multi-scale embedding, domain-related feature selection, similarity-contrastive and margin losses, and a weighted decision rule, evaluated on 12 industrial, medical, and video datasets.
Problem
Existing anomaly detection methods are domain-specific and generalize poorly across industrial, medical, and video domains, partly due to pre-trained feature biases and weak anomaly discrimination in supervised settings.
Approach
Extends ReContrast's student-teacher contrastive framework with a multi-scale embedding bottleneck, weighted domain-related feature selection guiding the student, a similarity-contrastive loss plus margin loss to separate normal and abnormal similarities, and a dynamic weighted top-k anomaly scoring rule, usable in both unsupervised and supervised settings.
Key results
- On MVTec AD, UniNet reaches 99.90% image-level AUROC and 96.00% PRO, improving the ReContrast baseline by 0.44% and 0.80%.
- On VAD (supervised), UniNet attains 99.95% I-AUROC without SegAD, versus 84.52% for ReContrast.
- On MVTec 3D-AD (RGB only), UniNet reaches 95.76% I-AUROC, versus 88.63% for ReContrast.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniNet instead of ReContrast | Unified anomaly detection · MVTec 3D-AD | I-AUROC: 88.63 -> 95.76 | Helped |
| Use UniNet instead of ReContrast | Unified anomaly detection · ISIC2018 | I-AUROC: 90.15 -> 100.0 | Helped |
| Use UniNet instead of ReContrast | Unified anomaly detection · APTOS | I-AUROC: 97.51 -> 100.0 | Helped |
| Use UniNet instead of ReContrast | Unified anomaly detection · OCT2017 | I-AUROC: 99.60 -> 100.0 | Helped |
| Use UniNet instead of ReContrast | Unified anomaly detection · Ped2 | I-AUROC: 95.2 -> 97.9 | Helped |
| Use UniNet instead of ReContrast | Unified anomaly detection · MVTec AD | P-AUROC: 99.03 -> 98.81 | Hurt |
| Add MEM (ablation) | UniNet · MVTec AD | I-AUROC: 98.42 -> 99.01 | Helped |
| Add DFS after MEM (ablation) | UniNet · MVTec AD | I-AUROC: 99.01 -> 99.23 | Helped |
| Use both LSC and LM rather than LSC alone (ablation) | UniNet · MVTec AD | I-AUROC: 99.36 -> 99.40 | Helped |
| Add the weighted decision mechanism (ablation) | UniNet · MVTec AD | I-AUROC: 99.77 -> 99.90 | Helped |
| Replace fixed 3% selection with the dynamic weighted decision mechanism (ablation) | UniNet · MVTec 3D-AD | I-AUROC: 94.42 -> 95.76 | Helped |
Surprising findings
- Unsupervised methods struggle on the VAD supervised benchmark without SegAD, and supervised methods also perform poorly due to unseen anomalies.
- Pre-trained features often bear little resemblance to the features needed for the target domain, adversely affecting performance.
Limitations
- Gains on some datasets are marginal or trail specific baselines (e.g., pixel AUROC below RealNet on MVTec AD; mIoU below MEGANet on Kvasir).
- Several supervised comparisons rely on a single benchmark setup and hyperparameters fixed across datasets.
Experimental setup
Datasets: MVTec AD BTAD MVTec 3D-AD VAD APTOS OCT2017 ISIC2018 Kvasir CVC-ClinicDB CVC-ColonDB Ped2
Benchmarks: MVTec AD BTAD MVTec 3D-AD VAD APTOS OCT2017 ISIC2018 Kvasir CVC-ClinicDB CVC-ColonDB Ped2
Metrics: I-AUROC P-AUROC PRO AP F1 ACC DSC mIoU
Baselines: ReContrast RD4AD PatchCore RealNet RD++ SimpleNet CFA AE-flow DevNet DRA EfficientAD MambaAD
Builds on: ReContrast Reverse distillation (RD4AD) SimCLR-style stop-gradient contrastive learning
Closest papers by meaning
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.83)
- Unsupervised Anomaly Detection Algorithms on Real-world Data: How Many Do We Need? (NeurIPS 2024 · similarity 0.82)
- Unsupervised Continual Anomaly Detection with Contrastively-Learned Prompt (AAAI 2024 · similarity 0.81)
- RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (CVPR 2024 · similarity 0.81)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.79)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.79)
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.