Research map / Label-efficient and robust learning / OOD and anomaly detection
WDiscOOD: Out-of-Distribution Detection via Whitened Linear Discriminant Analysis
Yiye Chen, Yunzhi Lin, Ruinian Xu, Patricio A. Vela
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 |
|---|---|---|---|
| combine discriminative and residual subspace scores (WDiscOOD) versus best baselines | WDiscOOD · ImageNet-1k OOD, ResNet-50 | AUROC: 86.64 -> 87.23 | Improved |
| apply WDiscOOD to contrastive encoders instead of Mahalanobis/KNN | SupCon and CLIP encoders · ImageNet-1k OOD | AUROC: SupCon 90.35 -> 90.89; CLIP 75.31 -> 75.74 | Improved |
| use whitened discriminative residual score instead of principal residual score (ablation) | Residual-subspace OOD score · ImageNet-1k OOD, ResNet-50 | AUROC: 84.01 -> 86.56 | Improved |
| add data whitening before LDA (ablation) | WDiscOOD · ImageNet-1k OOD, ResNet-50 | FPR95: 74.56 -> 49.86 | Improved |
| add data whitening before LDA (ablation) | WDiscOOD · ImageNet-1k OOD, ViT-B/16 | FPR95: 32.21 -> 26.49 | Improved |
| use Mahalanobis instead of Euclidean distance after whitened LDA (ablation) | WDiscOOD · ImageNet-1k OOD, ResNet-50 | FPR95: 49.86 -> 49.85 | No clear effect |
| add Mahalanobis distance after LDA without whitening (ablation) | WDiscOOD without whitening · ImageNet-1k OOD, ResNet-50 | FPR95: 74.56 -> 53.65 | Improved |
| combine WD and WDR subspace scores instead of a single subspace (ablation) | WDiscOOD · ImageNet-1k OOD | FPR95 | Improved |
| use extreme score scale factor alpha (very small or large) (ablation) | WDiscOOD · ImageNet-1k OOD, ResNet-50 | AUROC | Hurt |
| increase number of discriminants N_D (ablation) | WDiscOOD · ImageNet-1k OOD, ResNet-50 | AUROC | No clear effect |
| use 20% of training data (200K) for LDA statistics (ablation) | WDiscOOD · ImageNet-1k OOD, ResNet-50 | AUROC | No clear effect |
| apply feature-space methods on projection embedding instead of penultimate features (ablation) | Mahalanobis, KNN, WDiscOOD · SupCon and CLIP | FPR95: WDisc SupCon 40.10 -> 53.10; CLIP 77.57 -> 92.25 | Hurt |
Closest papers by meaning
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.89)
- Learning to Shape In-distribution Feature Space for Out-of-distribution Detection (NeurIPS 2024 · similarity 0.89)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.88)
- Exploiting Discrepancy in Feature Statistic for Out-of-Distribution Detection (AAAI 2024 · similarity 0.88)
- CASE: Exploiting Intra-class Compactness and Inter-class Separability of Feature Embeddings for Out-of-Distribution Detection (AAAI 2024 · similarity 0.87)
- Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 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.
