Research map / Detection and segmentation / Open-set and few-shot detection
Transfer-Based Semantic Anomaly Detection
Lucas Deecke, Lukas Ruff, Robert A. Vandermeulen, Hakan Bilen
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 L2 regularization toward pretrained ImageNet weights (inductive bias) to OE-based training (ablation) | ADIB · STL-10 semantic AD | mAP: 38.6 to 49.5 relative to pretrained no-reg (a2 vs ADIB) | Improved |
| use HSC loss with L2 regularization instead of eq.(2) objective (ablation) | semantic AD ablation · CIFAR-10 semantic AD | mAP: 68.5 (a3) vs 74.6 (ADIB) | Hurt |
| replace L2 regularization with EWC (ablation) | ADIB · CIFAR-10 semantic AD | mAP: 66.4 (a5) vs 74.6 (ADIB) | Hurt |
| use pretrained init without regularization (vs random init) (ablation) | semantic AD · CIFAR-10 semantic AD | mAP: 60.3 to 64.9 | Improved |
| use OE and transfer with residual adaptation (ADRA) instead of SAD baseline on MPI3D | ADRA · MPI3D interventions | AUC: ADRA improves performance over SAD under all interventions | Improved |
| add outlier exposure (OE) to DSVDD (ablation) | DSVDD · MPI3D color intervention | AUC: 11.7 to 100.0 | Improved |
| use SVHN instead of CIFAR-100 as OE dataset (ablation) | ADIB · CIFAR-10 one-versus-rest AD | mAUC: 99.1 with CIFAR-100 vs 79.8 with SVHN | Hurt |
| transfer-based regularization (ADIB/ADRA) vs SAD on small-mode robustness (ablation) | ADIB and ADRA · CIFAR-10 minor mode (ship/truck) | relative AUC for secondary class: SAD decreases much faster than ADIB and ADRA | Improved |
| adding pretrained-transfer inductive bias (ADRA/ADIB) to OE-based AD on CIFAR-10 one-vs-rest | ADIB · CIFAR-10 one-versus-rest AD | mAUC: 99.1 vs 96.3 (HSC) | Improved |
| add hyperspherical radial loss (HSC) vs standard classifier loss on CIFAR-10 semantic AD | HSC-based model · STL-10 semantic AD | mAP: 27.7 (HSC no reg) vs 32.9 (a1 standard loss) | No clear effect |
Closest papers by meaning
- Detecting Semantic Anomalies (AAAI 2020 · similarity 0.91)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.87)
- A Comprehensive Augmentation Framework for Anomaly Detection (AAAI 2024 · similarity 0.87)
- SetAD: Semi-Supervised Anomaly Learning in Contextual Sets (NeurIPS 2026 · similarity 0.87)
- A Dataset for Semantic Segmentation in the Presence of Unknowns (CVPR 2025 · similarity 0.86)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · 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.
