Research map / Label-efficient and robust learning / OOD and anomaly detection
TagFog: Textual Anchor Guidance and Fake Outlier Generation for Visual 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 TagFog, which trains an image encoder for visual OOD detection using Jigsaw-shuffled fake OOD images plus CLIP text-encoder anchors built from ChatGPT class descriptions, with contrastive losses, and reports state-of-the-art results on CIFAR10/100 and ImageNet100-I.
Problem
Visual out-of-distribution detection is hard when models are trained only on in-distribution data, leading to overconfidence on OOD inputs; obtaining real OOD training data is costly and existing fake-OOD methods (GANs, feature-space synthesis) are unstable or rely on strong distributional assumptio
Approach
Generate fake OOD images offline by Jigsaw-shuffling ID image patches, and guide the image encoder with contrastive alignment to CLIP text-encoder embeddings of ChatGPT-generated class descriptions (used as anchors), plus a supervised contrastive loss; the framework is then combined with post-hoc scores such as ReAct.
Key results
- On CIFAR100 with ResNet18, TagFog reaches average FPR95 44.33% and AUROC 90.47%, versus CIDER at 58.66% and 85.03%.
- On CIFAR10 with ResNet18, TagFog reaches average FPR95 10.41% and AUROC 97.99%, versus VOS at 33.04% and 92.99%.
- On ImageNet100-I with ResNet50, TagFog reaches average FPR95 29.50% and AUROC 94.67%, compared with ReAct at 39.85% and 92.12%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Jigsaw fake OOD data (ablation) | TagFog · CIFAR100, ResNet18 | AUROC: 87.43 -> 89.17 | Helped |
| add textual-anchor contrastive loss L_CI (ablation) | TagFog · CIFAR100, ResNet18 | AUROC: 87.43 -> 88.38 | Helped |
| add supervised contrastive loss L_SC (ablation) | TagFog · CIFAR100, ResNet18 | AUROC: 87.43 -> 88.27 | Helped |
| add all three components: fake OOD data, L_CI, and L_SC (ablation) | TagFog · CIFAR100, ResNet18 | AUROC: 87.43 -> 90.47 | Helped |
| replace ChatGPT class descriptions with “a photo of [ID class name]” (ablation) | TagFog textual-anchor guidance · CIFAR and ImageNet100 benchmarks | OOD detection performance | Hurt |
| replace CLIP textual anchors with random embeddings (ablation) | TagFog textual-anchor guidance · CIFAR and ImageNet100 benchmarks | OOD detection performance | Hurt |
Surprising findings
- Replacing ChatGPT descriptions with a simple 'a photo of [class]' prompt or random anchors degrades OOD performance, indicating the richer descriptions matter.
- The Jigsaw fake OOD data alone, without text guidance or SupCon, already improves CIFAR100 average AUROC from 87.43% to 89.17%.
Limitations
- Relies on ChatGPT-generated descriptions, whose quality and generation are not systematically controlled.
- Gains on some OOD datasets (e.g., Textures, Places365 on CIFAR100) are weaker than specific baselines such as FeatureNorm or CIDER.
- Hyperparameter sensitivity is evaluated mainly on CIFAR100 with ResNet18.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet100-I ImageNet100-II Textures SVHN iSUN Places365 LSUN-C LSUN-R iNaturalist SUN
Benchmarks: CIFAR10 CIFAR100 ImageNet100-I ImageNet100-II
Metrics: FPR95 AUROC
Baselines: MSP Mahalanobis ODIN Energy ViM DICE BATS ReAct FeatureNorm LINe VOS LogitNorm
Builds on: CLIP SupCon ReAct VOS CIDER
Closest papers by meaning
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.89)
- Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection (ICML 2024 · similarity 0.88)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.87)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.87)
- Unsupervised Layer-Wise Score Aggregation for Textual OOD Detection (AAAI 2024 · similarity 0.87)
- Diffusion-based Semantic Outlier Generation via Nuisance Awareness for Out-of-Distribution Detection (AAAI 2025 · 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.