Research map / Label-efficient and robust learning / OOD and anomaly detection
OODD: Test-time Out-of-Distribution Detection with Dynamic Dictionary
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 OODD, a test-time OOD detector that maintains a priority-queue OOD dictionary of low-scoring test features (plus ID-derived outliers for stabilization) and calibrates KNN-style scores without fine-tuning.
Problem
Out-of-distribution detection for deep classifiers when test-time OOD samples differ from training outliers; existing test-time calibration methods suffer from catastrophic forgetting and unstable detection from batch-wise parameter updates.
Approach
Replace parameter updates with a dynamic, priority-queue-managed dictionary of latent OOD features collected during testing, and use it to calibrate KNN-based scores; initialize it with outliers cropped from training ID data (C-Out) and sample informative inliers via patch/class-level confidence filtering; cosine similarity replaces Euclidean distance for speed.
Key results
- On CIFAR-100 Far OOD, OODD reaches 24.74% FPR95 vs. 58.17% for TULIP (the abstract reports a 26.0% improvement over the state of the art).
- On CIFAR-10 Far OOD, OODD achieves 17.44% FPR95 vs. 24.27% for KNN.
- Time cost on CIFAR-10 is 3.98 s vs. 9.61 s for KNN, roughly a 3x speedup over KNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the dynamic OOD dictionary with informative inlier sampling | KNN OOD detection · CIFAR-100 Far OOD | FPR95: 53.65 -> 24.74 | Helped |
| add the dynamic OOD dictionary with informative inlier sampling | KNN OOD detection · ImageNet-200 Far OOD | AUROC: 93.16 ± 0.22 -> 95.47 ± 0.03 | Helped |
| add informative inlier sampling (ablation) | OODD on ImageNet-200 · Near OOD | FPR95: 60.57 -> 58.92 | Helped |
| add informative inlier sampling (ablation) | OODD on ImageNet-200 · Far OOD | AUROC: 91.47 -> 92.83 | Helped |
| add the priority queue (ablation) | OODD on ImageNet-200 · Near OOD | FPR95: 60.57 -> 53.85 | Helped |
| add the priority queue (ablation) | OODD on ImageNet-200 · Far OOD | AUROC: 91.47 -> 95.45 | Helped |
| add the memory bank (ablation) | OODD on ImageNet-200 · Near OOD | FPR95: 60.57 -> 59.14 | Helped |
| add the memory bank (ablation) | OODD on ImageNet-200 · Far OOD | AUROC: 91.47 -> 88.89 | Hurt |
| increase priority queue size from 512 to 20k (ablation) | OODD on ImageNet-200 · Near OOD | AUROC: 84.82 -> 83.21 | Hurt |
| increase priority queue size from 512 to 20k (ablation) | OODD on ImageNet-200 · Far OOD | AUROC: 95.15 -> 89.26 | Hurt |
| initialize the OOD dictionary with C-Out instead of no outliers (ablation) | OODD on ImageNet-200 · Near OOD | AUROC: 85.74 vs. 85.69 | No clear effect |
| replace Euclidean distance with cosine similarity | KNN-based OOD detection · CIFAR-10 and ImageNet-200 | speedup: 3x | Helped |
Surprising findings
- Initializing the OOD dictionary empty has little impact on final detection in most scenarios, though early-stage AUROC fluctuates more without initialization.
- Adding the OOD dictionary to logit-based methods (MaxLogits, Energy) gives only modest gains, while KNN benefits substantially.
Limitations
- Gains on Near OOD are small or absent (e.g., CIFAR-100 near FPR95 not best), and on ImageNet-1K Near OOD OODD does not lead.
- Gains for logit-based post-hoc methods are modest because the dictionary stores latent features.
- Performance depends on hyperparameters such as queue size and K, which are tuned on OpenOOD validation sets.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-200 ImageNet-1K MNIST SVHN Textures Places365 Tiny ImageNet-200 SSB-hard NINCO iNaturalist
Benchmarks: OpenOOD CIFAR-10 OpenOOD CIFAR-100 OpenOOD ImageNet-200 OpenOOD ImageNet-1K
Metrics: FPR95 AUROC ID ACC
Baselines: MaxLogits ODIN Mahalanobis distance Energy KNN ViM RTL++ TULIP
Builds on: KNN OOD detection TULIP RTL++ OpenOOD
Closest papers by meaning
- Test-Time Linear Out-of-Distribution Detection (CVPR 2024 · similarity 0.91)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.89)
- What If the Input is Expanded in OOD Detection? (NeurIPS 2024 · similarity 0.89)
- Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.88)
- Expecting The Unexpected: Towards Broad Out-Of-Distribution Detection (NeurIPS 2024 · similarity 0.88)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.88)
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.