Research map / Label-efficient and robust learning / OOD and anomaly detection
Adaptive Multi-prompt Contrastive Network 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 AMCN, a CLIP-based few-shot OOD detection method that learns ID prompts plus fixed and adaptive OOD prompts with class-wise adaptive thresholds and margin losses.
Problem
Out-of-distribution detection when only a few labeled ID images per class are available, with no OOD training data, so that per-class diversity and background bias hurt ID-OOD boundaries.
Approach
Learns three kinds of text prompts (learnable ID prompts, label-fixed OOD prompts from disjoint dataset labels, and label-adaptive OOD prompts with learnable labels) on a frozen CLIP, adds a class-wise adaptive threshold from per-class mean and standard deviation of scores, and enforces a prompt-level ID-OOD margin via contrastive and alignment losses.
Key results
- On Texture with one shot, AMCN reaches FPR95 39.16 and AUROC 89.88, versus 48.87 FPR95 for SCT (the reported 9.71-point FPR95 gain).
- Averaged over four OOD sets, AMCN achieves 30.56 FPR95 and 94.29 AUROC at eight shots, versus 32.32 and 93.53 for SCT.
- Ablation on iNaturalist shows the adaptive threshold improves eight-shot FPR95 from 20.51 (fixed) to 18.17.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AMCN instead of the strongest reported baseline SCT | few-shot OOD detection · Texture (1-shot) | FPR95: 48.87 -> 39.16 | Helped |
| use AMCN instead of the strongest reported baseline SCT | few-shot OOD detection · Places (1-shot) | AUROC: 91.23 -> 92.78 | Helped |
| use AMCN instead of the strongest reported baseline SCT | few-shot OOD detection · SUN (1-shot) | AUROC: +0.27% | Helped |
| use AMCN instead of the strongest reported baseline SCT | few-shot OOD detection · SUN (8-shot) | AUROC: +1.12% | Helped |
| remove Adaptive Prompt Generation (ablation) | AMCN · Texture (one-shot) | AUROC: 89.88 -> 82.59 | Hurt |
| remove Adaptive Prompt Generation (ablation) | AMCN · Texture (eight-shot) | AUROC: 93.43 -> 86.24 | Hurt |
| remove Prompt-based Multi-diversity Distribution Learning (ablation) | AMCN · Texture (one-shot) | AUROC: 89.88 -> 83.24 | Hurt |
| remove Prompt-based Multi-diversity Distribution Learning (ablation) | AMCN · Texture (eight-shot) | AUROC: 93.43 -> 86.88 | Hurt |
| remove Prompt-guided OOD Detection (ablation) | AMCN · Texture (one-shot) | AUROC: 89.88 -> 83.19 | Hurt |
| remove Prompt-guided OOD Detection (ablation) | AMCN · Texture (eight-shot) | AUROC: 93.43 -> 87.92 | Hurt |
| replace the adaptive class-wise threshold with a fixed threshold (ablation) | AMCN · iNaturalist (one-shot) | AUROC: 96.18 -> 93.83 | Hurt |
| replace the adaptive class-wise threshold with a fixed threshold (ablation) | AMCN · iNaturalist (eight-shot) | AUROC: 96.89 -> 94.16 | Hurt |
Surprising findings
- LoCoOp and CoOp in several cases perform worse than the zero-shot GL-MCM baseline, which the authors attribute to background bias.
Limitations
- Few-shot OOD results rely on OOD label names from other datasets that are disjoint from ID classes, so OOD prompts depend on available label knowledge.
- Many hyperparameters (epsilon, tau, lambda, alpha, S, Z) are tuned and the paper notes sensitivity to distribution margins.
Experimental setup
Datasets: ImageNet-1k iNaturalist Places Texture SUN CIFAR-100
Benchmarks: Texture Places SUN iNaturalist
Metrics: FPR95 AUROC
Baselines: KNN ViM ODIN NPOS GL-MCM MCM SeTAR CoOp LoCoOp SCT
Builds on: CoOp LoCoOp MCM CLIP
Closest papers by meaning
- ID-like Prompt Learning for Few-Shot Out-of-Distribution Detection (CVPR 2024 · similarity 0.92)
- Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection (ICLR 2025 · similarity 0.90)
- MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.89)
- Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection (ICML 2024 · similarity 0.89)
- GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.88)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 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.