Research map / Label-efficient and robust learning / OOD and anomaly detection
AdaNeg: Adaptive Negative Proxy Guided OOD Detection with Vision-Language Models
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 AdaNeg, a training-free, annotation-free OOD detector for CLIP that builds a test-time feature memory of confidently classified OOD and ID images and derives task-adaptive and sample-adaptive negative proxies to complement static NegLabel negative text labels.
Problem
Zero-shot OOD detection with vision-language models using fixed negative text labels suffers semantic misalignment with the actual OOD label space of each target dataset.
Approach
Dynamically generate negative proxies during testing by caching discriminative test image features into a category-split memory bank; average them (task-adaptive) or attention-weight them per test sample (sample-adaptive), and fuse the resulting scores with the static NegLabel score.
Key results
- On ImageNet with four OOD datasets (ViT-B/16), AdaNeg reaches 96.66 average AUROC and 18.92 FPR95, versus NegLabel's 94.21 and 25.40.
- On the OpenOOD ImageNet setup, AdaNeg improves near-OOD FPR95 to 67.51 and far-OOD FPR95 to 17.31 relative to NegLabel's 69.45 and 23.73.
- Training-free AdaNeg reports 476 FPS versus NegLabel's 592 FPS, with no learnable parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace fixed negative-label proxies with adaptive proxies | NegLabel on ImageNet-1k OOD average · iNaturalist, SUN, Places, Textures | AUROC: 94.21 -> 96.66 | Helped |
| replace fixed negative-label proxies with adaptive proxies | NegLabel on ImageNet-1k OOD average · iNaturalist, SUN, Places, Textures | FPR95: 25.40 -> 18.92 | Helped |
| replace fixed negative-label proxies with adaptive proxies | NegLabel on ImageNet-1k OpenOOD · near-OOD and far-OOD | AUROC: 75.18 -> 76.70; 94.85 -> 96.43 | Helped |
| add task-adaptive proxies (ablation) | NegLabel score on ImageNet-1k OpenOOD · near-OOD and far-OOD | AUROC: 75.18 -> 76.49; 94.85 -> 96.45 | Helped |
| add sample-adaptive proxies (ablation) | NegLabel score on ImageNet-1k OpenOOD · near-OOD and far-OOD | AUROC: 75.18 -> 76.70; 94.85 -> 96.63 | Helped |
| use sample-adaptive rather than task-adaptive proxies alone (ablation) | Adaptive proxy score on ImageNet-1k OpenOOD · near-OOD and far-OOD | AUROC: 75.76 -> 76.03; 96.20 -> 96.35 | Helped |
| combine adaptive image-based proxies with fixed text-based proxies (ablation) | AdaNeg on ImageNet-1k OpenOOD · near-OOD and far-OOD | AUROC: 76.03 -> 76.70; 96.35 -> 96.63 | Helped |
| adaptively adjust the memory gap using the estimated ID/OOD ratio (ablation) | AdaNeg · ImageNet and SUN, ID:OOD ratios 1:100, 1:10, 1:1, 10:1, 100:1, 1K:1 | FPR95: 21.00 -> 20.50; 12.49 -> 12.22; 9.81 -> 9.73; 15.61 -> 12.98; 20.71 -> 15.61; 26.28 -> 18.43 | Helped |
| use AdaNeg instead of NegLabel | OOD detection on ImageNet and SUN · ID:OOD ratio 1K:1 | FPR95: 23.69 -> 26.28 | Hurt |
| increase feature memory length (ablation) | AdaNeg · ImageNet-1k OpenOOD | OOD detection performance | Helped |
| vary the score-combination weight λ (ablation) | AdaNeg · ImageNet-1k OpenOOD | OOD detection performance | No clear effect |
| vary the sample-attention sharpness β (ablation) | AdaNeg · ImageNet-1k OpenOOD | OOD detection performance | No clear effect |
Surprising findings
- Under highly imbalanced ID:OOD test mixtures (1000:1), AdaNeg has a negative impact versus NegLabel, which the authors address with an adaptive gap (AdaGap).
Limitations
- Requires external memory storage (about 214.75MB for ImageNet ID), which may be an issue for storage-constrained settings.
- Performance degrades under extremely unbalanced ID:OOD test ratios (1000:1) without the AdaGap fix.
- Gap hyperparameter g works best with different values for near- versus far-OOD, so the default is a compromise.
Experimental setup
Datasets: ImageNet-1k NegLabel negative label corpus SSB-hard NINCO iNaturalist SUN Places Textures OpenImage-O CIFAR10 CIFAR100 TIN
Benchmarks: ImageNet-1k OOD (iNaturalist, SUN, Places, Textures) OpenOOD ImageNet-1k near/far-OOD OpenOOD CIFAR10/CIFAR100 BIMCV-COVID19+ medical OOD
Metrics: AUROC FPR95 ID ACC
Baselines: NegLabel MCM LAPT NegPrompt LoCoOp CLIPN LSN ZOC Mahalanobis Energy
Builds on: NegLabel CLIP MCM
Closest papers by meaning
- Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Negative Label Guided OOD Detection with Pretrained Vision-Language Models (ICLR 2024 · similarity 0.89)
- Enhanced OoD Detection through Cross-Modal Alignment of Multi-Modal Representations (CVPR 2025 · similarity 0.88)
- Reflexive Guidance: Improving OoDD in Vision-Language Models via Self-Guided Image-Adaptive Concept Generation (ICLR 2025 · similarity 0.88)
- GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.87)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.87)
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.