Research map / Label-efficient and robust learning / OOD and anomaly detection
Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection
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 Local-Prompt, a few-shot tuning method that freezes global CLIP prompts and learns extra local prompts with random-crop negative augmentation and regional regularization to improve OOD detection.
Problem
Few-shot OOD detection with vision-language models relies on global prompts and ignores local regional information where hard outliers differ only subtly in certain image regions.
Approach
Freeze global hand-crafted prompts to guide random-crop negative augmentation, and train local prompts (plus local negative prompts) with a top-k regional contrastive loss, diversity regularization, and a regional OOD score; the local prompts can be combined with trained global prompts at inference.
Key results
- 4-shot Local-Prompt reduces average FPR95 on ImageNet-1k to 27.66 vs 33.82 for CoOp (4-shot) and improves average AUROC to 93.53.
- Local-Prompt+LoCoOp at 4-shot achieves average FPR95 26.29 and AUROC 94.03, surpassing 16-shot LoCoOp (28.66 FPR95).
- ID accuracy on ImageNet-1k at 16-shot is 74.24 for Local-Prompt vs 71.86 for LoCoOp.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add local negative loss to positive local loss (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | FPR95: 30.85 -> 29.27 | Helped |
| Add diversity regularization to positive and negative local losses (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | AUROC: 93.27 -> 93.48 | Helped |
| Replace random crop with cutout augmentation (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | FPR95: 28.77 -> 33.21 | Hurt |
| Replace random crop with Gaussian noise augmentation (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | FPR95: 28.77 -> 30.76 | Hurt |
| Use 100 rather than 300 local negative prompts (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | FPR95: 28.77 -> 28.82 | Hurt |
| Use 200 rather than 300 local negative prompts (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | FPR95: 28.77 -> 28.52 | Helped |
| Increase negative augmentations from 1 to 8 (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | FPR95: 28.77 -> 30.04 | Hurt |
| Train for 50 rather than 30 epochs (ablation) | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | AUROC: 93.48 -> 93.40 | Hurt |
| Integrate trained LoCoOp global prompts with local prompts | Local-Prompt · ImageNet-1k ID, 4-shot OOD average | FPR95: 27.66 -> 26.29 | Helped |
| Use local prompts instead of LoCoOp | CLIP-Base/16 · ImageNet-1k ID, 4-shot OOD average | FPR95: 31.46 -> 27.66 | Helped |
| Use local prompts instead of LoCoOp | CLIP-Base/16 · ImageNet-1k ID, 4-shot | ID accuracy: 69.50 -> 70.52 | Helped |
| Use Local-Prompt for near-OOD detection | CLIP-Base/16 · ImageNet-10 ID, ImageNet-20 OOD, 16-shot | FPR95: 11.20 -> 3.90 | Helped |
Surprising findings
- Local-Prompt without any real outliers outperforms NegLabel (which uses outlier corpus labels) on average FPR95 (24.52 vs 25.40).
- Random crop augmentation was the only augmentation that helped; cutout and Gaussian noise degraded performance relative to the baseline.
Limitations
- Extra training time due to negative augmentation forward passes.
- Only applied to image classification OOD detection; not extended to segmentation or detection.
- Global prompts are kept frozen with a hand-crafted template in the main method, leaving joint global-local training unexplored.
Experimental setup
Datasets: ImageNet-1k ImageNet-100 ImageNet-10 ImageNet-20 iNaturalist SUN Places Texture
Benchmarks: ImageNet-1k OOD (iNaturalist, SUN, Places, Texture) ImageNet-100 OOD ImageNet-10 vs ImageNet-20 near OOD ImageNet-1k ID accuracy
Metrics: FPR95 AUROC ID accuracy
Baselines: MCM GL-MCM CoOp LoCoOp NegLabel ID-like MSP Energy KNN NPOS Textual-OE
Builds on: CoOp LoCoOp MCM GL-MCM
Closest papers by meaning
- Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection (ICML 2025 · similarity 0.90)
- Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection (ICML 2024 · similarity 0.89)
- ID-like Prompt Learning for Few-Shot Out-of-Distribution Detection (CVPR 2024 · similarity 0.88)
- Self-Calibrated Tuning of Vision-Language Models for Out-of-Distribution Detection (NeurIPS 2024 · similarity 0.87)
- Overcoming Shortcut Problem in VLM for Robust Out-of-Distribution Detection (CVPR 2025 · similarity 0.87)
- DeCoOp: Robust Prompt Tuning with Out-of-Distribution Detection (ICML 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.