Research map / Vision-language and multimodal / CLIP and zero-shot
Category-Extensible Out-of-Distribution Detection via Hierarchical Context Descriptions
Kai Liu, Zhihang Fu, Chao Chen, Sheng Jin, Ze Chen, Mingyuan Tao, Rongxin Jiang, Jieping Ye
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 |
|---|---|---|---|
| Learn hierarchical perceptual and spurious contexts instead of using a perceptual-only baseline (ablation) | CATEX on ImageNet-100 · ImageNet-100 | FPR95: 13.07 -> 11.39 | Improved |
| Learn hierarchical contexts using perturbation-guided synthetic OOD samples (ablation) | CATEX on ImageNet-100 · ImageNet-100 | FPR95: 11.39 -> 10.59 | Improved |
| Regularize inference similarities with the spurious-context score (ablation) | CATEX on ImageNet-100 · ImageNet-100 | FPR95: 10.59 -> 10.31 | Improved |
| Use a larger perturbing ratio, especially perturbing all context tokens (ablation) | CATEX · ImageNet-100 | FPR95: 10.31 -> 11.70 | Hurt |
| Choose the most distant token for perturbation rather than a random token (ablation) | CATEX · ImageNet-100 | FPR95: 10.31 -> 10.42 | Hurt |
| Choose the closest token for perturbation rather than a random token (ablation) | CATEX · ImageNet-100 | FPR95: 10.31 -> 10.86 | Hurt |
| Add an orthogonal constraint when using multiple spurious contexts (ablation) | CATEX with 4 spurious contexts · ImageNet-100 | FPR95: 10.27 -> 9.89 | Improved |
| Increase the number of unconstrained spurious contexts (ablation) | CATEX · ImageNet-100 | FPR95: 10.31 -> 10.21 | No clear effect |
| Use the regularized score with simulated spurious contexts rather than visual descriptions alone | Zero-shot classification with CLIP and VisDesc · ImageNet-1K zero-shot classification | ACC-Top1: 65.47 -> 65.84 | Improved |
| Use CATEX rather than NPOS for standard OOD detection | ImageNet-1K ID, average over OOD datasets · ImageNet-1K ID; OOD average | FPR95: 37.93 -> 29.66 | Improved |
| Merge independently learned contexts for the two category subsets | CATEX · ImageNet-200 union | ACC: 86.23 -> 89.61 | Improved |
| Use CATEX instead of MCM under a severe domain shift | ImageNet-1K-trained model · ImageNet-Sketch | ACC: 47.42 -> 48.59 | Hurt |
Closest papers by meaning
- Out-of-Distribution Detection with Positive and Negative Prompt Supervision Using Large Language Models (AAAI 2026 · similarity 0.91)
- Just a Few Glances: Open-Set Visual Perception with Image Prompt Paradigm (AAAI 2025 · similarity 0.90)
- Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- RUNA: Object-Level Out-of-Distribution Detection via Regional Uncertainty Alignment of Multimodal Representations (AAAI 2025 · similarity 0.88)
- DiPEx: Dispersing Prompt Expansion for Class-Agnostic Object Detection (NeurIPS 2024 · similarity 0.88)
- Learning Background Prompts to Discover Implicit Knowledge for Open Vocabulary Object Detection (CVPR 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.
