Research map / Vision-language and multimodal / Human and embodied modeling
On the Powerfulness of Textual Outlier Exposure for Visual OoD Detection
Sangha Park, Jisoo Mok, Dahuin Jung, Saehyung Lee, Sungroh Yoon
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 |
|---|---|---|---|
| Use caption-level textual outliers instead of no outlier exposure | CLIP classifier · ImageNet-1K OoD average | FPR95: 93.87 -> 58.21 | Improved |
| Use caption-level textual outliers instead of VOS visual outliers | OoD detection · ImageNet-1K OoD average | FPR95: 87.87 -> 58.21 | Improved |
| Use textual outliers rather than visual auxiliary images | Outlier exposure · ImageNet20 ID vs ImageNet10 OoD | AUROC: 98.22 ± 0.58 vs 98.14 ± 0.17 | Mixed |
| Add prompt ensembling to word-level textual outliers (ablation) | Word outlier exposure · ImageNet-1K OoD average | FPR95: 64.73 -> 59.11 | Improved |
| Add prompt ensembling to description-level textual outliers (ablation) | Description outlier exposure · ImageNet-1K OoD average | FPR95: 70.60 -> 65.40 | Improved |
| Add Gaussian noise to textual embeddings (ablation) | Caption outlier exposure · ImageNet-1K OoD average | FPR95: 64.14 -> 58.35 | Improved |
| Add Gaussian noise to textual embeddings (ablation) | Caption outlier exposure · ImageNet-1K OoD average | AUROC: 83.26 -> 85.53 | Improved |
| Generate textual outliers from captions rather than using image embeddings directly (ablation) | Caption-level outlier exposure · OoD datasets evaluated in the ablation | AUROC | Improved |
| Use a larger CLIP model (ablation) | Textual outlier exposure · ImageNet-1K OoD detection | AUROC | Improved |
| Use description-level textual outliers instead of word-level outliers | Hard OoD detection · CUB100 hard OoD | AUROC: 71.62 -> 72.33 | Improved |
| Use description-level textual outliers instead of caption-level outliers | Hard OoD detection · CUB100 hard OoD | AUROC: 64.79 -> 72.33 | Improved |
Closest papers by meaning
- Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection (ICML 2024 · similarity 0.91)
- RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples (ICML 2024 · similarity 0.90)
- Unsupervised Layer-Wise Score Aggregation for Textual OOD Detection (AAAI 2024 · similarity 0.89)
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.89)
- Local-Prompt: Extensible Local Prompts for Few-Shot Out-of-Distribution Detection (ICLR 2025 · similarity 0.89)
- Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection (ICML 2025 · 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.
