Research map / Label-efficient and robust learning / OOD and anomaly detection
If It's Not Efficient, It's Not Usable: Real-Time OOD Detection with Latent De-Biasing and High-Quality Negative Samples
Yuchuan Li, Jae-Mo Kang, Il-Min Kim
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 latent-space pixel-intensity de-biasing (32 perturbed variants) (ablation) | VAE OOD detector, no negatives · CIFAR-100 ID, Tiny-ImageNet | AUROC: 47.5 -> 58.0 | Improved |
| add latent-space pixel-intensity de-biasing (32 perturbed variants) (ablation) | VAE OOD detector, no negatives · ImageNet-1k ID, NINCO | AUROC: 43.5 -> 57.0 | Improved |
| increase perturbed variants from 4 to 32 (ablation) | VAE OOD detector · CIFAR-100 ID, SVHN | AUROC: 56.9 -> 63.7 | Improved |
| add Mosaic-based initialization of negatives (step 1) (ablation) | Base VAE · CIFAR-100 ID, SVHN | AUROC: 63.7 -> 67.6 | Improved |
| add iterative negative update (step 2) (ablation) | Base + s1 · CIFAR-100 ID, SVHN | AUROC: 67.6 -> 71.4 | Improved |
| add screening of easy and invalid negatives (step 3) (ablation) | Base + s1 + s2 · CIFAR-100 ID, SVHN | AUROC: 71.4 -> 75.5 | Improved |
| add screening of negatives (step 3) (ablation) | Base + s1 + s2 · ImageNet-1k ID, NINCO | AUROC: 60.2 -> 62.0 | Improved |
| use proposed negative generation instead of HamOS (ablation) | Stage 2 replacement · CIFAR-100 ID, Tiny-ImageNet | AUROC: 66.0 -> 67.3 | Improved |
| use proposed negative generation instead of CSI (ablation) | Stage 2 replacement · ImageNet-1k ID, NINCO | AUROC: 53.4 -> 62.0 | Improved |
| use proposed method instead of strongest VAE baseline DP | TinyML latency-matched OOD detectors · CIFAR-100 ID, Tiny-ImageNet | AUROC: 65.9 -> 67.3 | Improved |
| replace importance-sampling de-biasing with latent de-biasing | VAE OOD detector · ImageNet-1k, Jetson latency | latency (ms): 1.27 -> 0.12 ms | Improved |
Closest papers by meaning
- Fast Decision Boundary based Out-of-Distribution Detector (ICML 2024 · similarity 0.88)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.87)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.87)
- Can We Ignore Labels in Out of Distribution Detection? (ICLR 2025 · similarity 0.86)
- MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.86)
- OODD: Test-time Out-of-Distribution Detection with Dynamic Dictionary (CVPR 2025 · 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.
