Research map / Label-efficient and robust learning / OOD and anomaly detection
Mean-Shifted Contrastive Loss for Anomaly Detection
Tal Reiss, Yedid Hoshen
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 |
|---|---|---|---|
| Fine-tune ImageNet-pretrained features with mean-shifted contrastive loss instead of PANDA center loss | ResNet152 anomaly detection · CIFAR-10 | ROC-AUC: 96.2 -> 97.2 | Improved |
| Replace standard contrastive loss with mean-shifted contrastive loss | ImageNet-pretrained feature adaptation · One-class anomaly detection | ROC-AUC | Improved |
| Add angular center loss to MSC (ablation) | ResNet152 anomaly detection · CIFAR-10 | ROC-AUC: 97.2 -> 97.5 | Improved |
| Replace Euclidean center loss with angular center loss (ablation) | PANDA · CIFAR-10 | ROC-AUC: 96.2 -> 96.8 | Improved |
| Normalize pretrained features to remove confidence magnitude (ablation) | ImageNet-pretrained feature anomaly detection · CIFAR-10 | ROC-AUC: 92.5 -> 95.8 | Improved |
| Use mean-shifted contrastive loss across different pretrained backbones | ImageNet-pretrained anomaly detection · CIFAR-10 | ROC-AUC: 89.3 -> 97.0; 85.6 -> 95.7; 95.7 -> 98.6 | Improved |
| Initialize MSC without a strong pretrained feature prior (ablation) | MSC feature adaptation · Anomaly detection | ROC-AUC | Hurt |
| Use ImageNet-pretrained features with CSI rather than the standard CSI setup (ablation) | CSI · CIFAR-10 | ROC-AUC: 94.3 -> 89.5 | Hurt |
| Increase the CSI backbone size from ResNet18 to larger backbones (ablation) | CSI · CIFAR-10 | ROC-AUC: 94.3% | No clear effect |
| Compress the training gallery with k-means instead of using the full training set (ablation) | MSC anomaly scoring · CIFAR-10 OCC | ROC-AUC: 97.2 -> 94.2; 97.2 -> 95.8; 97.2 -> 96.1; 97.2 -> 97.0 | Mixed |
| Train standard contrastive loss with smaller temperature (ablation) | ImageNet-pretrained contrastive adaptation · CIFAR-10 | ROC-AUC | Hurt |
| Use the mean-shifted objective on multi-class normal data | Pretrained feature adaptation · CIFAR-10 multi-class AD | ROC-AUC: 79.0 -> 85.3 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PANDA — PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation (CVPR 2021) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Supervised Anomaly Detection With Neural Transformations (TPAMI 2024 · similarity 0.89)
- CSI: Novelty Detection via Contrastive Learning on Distributionally Shifted Instances (NeurIPS 2020 · similarity 0.88)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.87)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.86)
- RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (CVPR 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.
