Research map / Detection and segmentation / Open-set and few-shot detection
Fake It Until You Make It : Towards Accurate Near-Distribution Novelty Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 pretrained features with synthetic near-anomalies generated by SDE (ablation) | ViT-B_16 novelty detector · CIFAR-10vs100 | AUROC (%): 65.5 -> 90.0 | Helped |
| Fine-tune pretrained features with synthetic near-anomalies generated by SDE (ablation) | ViT-B_16 novelty detector · CIFAR-10 | AUROC (%): 91.0 -> 99.1 | Helped |
| Use SDE-generated fake anomalies rather than GAN-generated anomalies (ablation) | ViT-B_16 novelty detector · Weather | AUROC (%): 82.4 -> 97.0 | Helped |
| Use SDE-generated fake anomalies rather than GAN-generated anomalies (ablation) | ViT-B_16 novelty detector · CIFAR-10 | AUROC (%): 97.8 -> 99.1 | Helped |
| Replace SDE fake samples with other generative-model samples (ablation) | ViT-B_16 novelty detector · Weather and MvTecAD | AUROC (%): Weather: 69.6, 83.7, 85.3, 82.4, 76.8, 84.9; MvTecAD: 68.8, 71.8, 76.4, 80.5, 73.2, 78.3 | Hurt |
| Vary the feature-extractor backbone while retaining generated-data fine-tuning (ablation) | Proposed detector · CIFAR-10vs100 | AUROC (%): ResNet-152: 58.5 -> 74.7; ViT-B_16: 65.5 -> 90.0; ViT-B_32: 66.3 -> 78.9; R50+ViT-B_16: 50.0 -> 85.9; ConvNeXt-B: 71.5 - | Mixed |
| Change k in the k-nearest-neighbor anomaly score (ablation) | Proposed detector · CIFAR-10vs100 | AUROC (%): 89.8, 90.0, 90.0, 90.0, 89.7 | No clear effect |
| Change the SDE stopping point to FID ranges outside 30<FID<50 (ablation) | Proposed detector · CIFAR-10vs100 | AUROC (%): 82.7, 87.1, 90.0, 68.2 | Hurt |
| Use the closest CLP anomaly class instead of the bottom-1 anomaly class (ablation) | Ours · CIFAR-10 near-ND | AUROC (%): 90.0 -> 92.8 | Helped |
| Use near-ND-trained method on near anomalies rather than standard far anomalies | Existing novelty detectors · Near-ND versus regular ND | AUROC (%): up to 20% decrease | Hurt |
Closest papers by meaning
- Generative Probabilistic Novelty Detection with Adversarial Autoencoders (NeurIPS 2018 · similarity 0.89)
- Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability (CVPR 2024 · similarity 0.87)
- Universal Novelty Detection Through Adaptive Contrastive Learning (CVPR 2024 · similarity 0.86)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.86)
- Secret Lies in Color: Enhancing AI-Generated Images Detection with Color Distribution Analysis (CVPR 2025 · similarity 0.85)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.85)
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.