Abstract
In safety-critical domains such as autonomous driving and medical diagnostics, vision systems must not only be accurate but also self-aware. When encountering an unfamiliar "outlier," a model should flag it rather than making an erroneous high-confidence prediction. We introduce UNO (Uncertainty + Negative Objectness), a method that improves outlier detection by decoupling in-distribution uncertainty from explicit outlier recognition. By ensembling these two distinct signals, our approach achieves state-of-the-art performance in both image classification and complex pixel-level segmentation tasks.
Introduction
Most contemporary outlier detection methods attempt to force models to produce low-confidence predictions for unfamiliar data. This is typically achieved by training on a "negative" dataset—images the model should not recognize—and penalizing high confidence. However, this approach conflates two distinct factors: in-distribution uncertainty (confusion between known classes) and outlier recognition (identifying data that resembles known negatives). Conflating these signals can lead to failures in detecting outliers that differ from training negatives or produce false positives at the boundaries of known classes.
Methodology: The UNO Framework
The core of the proposed method is the UNO score. Rather than a standard $K$-way classifier, the authors utilize a $K+1$ classifier, where the $(K+1)$-th logit is dedicated to a "negative" class.
The UNO Score
The outlier score is defined as the sum of two decoupled components:
- $s_{Unc}$ (Uncertainty): Calculated as the negative maximum confidence across the $K$ known classes. This component identifies "confused" predictions where the model cannot clearly distinguish between inliers.
- $s_{NO}$ (Negative Objectness): The posterior probability of the $(K+1)$-th class. This component identifies samples that specifically resemble the negative data seen during training.
Geometric and Dense Adaptations
Geometrically, class weight vectors in the pre-logit feature space tend toward orthogonality. Uncertainty is triggered by small feature norms or poor alignment with inlier vectors, while Negative Objectness is triggered by large feature norms aligning with the $(K+1)$-th weight vector.
For dense prediction tasks like road scene segmentation, the authors adapt the Mask2Former architecture. They extend its taxonomy to $K+2$ classes: $K$ inliers, one "negative objectness" class (trained on outliers), and one "no-object" class (handling unused queries). This distinction is crucial, as unused queries (representing "nothing") differ significantly from actual outliers (representing "something unknown").
Key Insights and Results
A primary finding is that $s_{Unc}$ and $s_{NO}$ are largely uncorrelated. Their distinct failure modes make them ideal for ensembling.
Performance Benchmarks
The authors evaluated UNO across several rigorous benchmarks:
- Segmentation: On the Fishyscapes "Static" benchmark, UNO achieved an Average Precision (AP) of 95.8%, a significant improvement over the previous state-of-the-art (92.4%).
- Classification: On OpenOOD CIFAR-10, UNO reached a Near-OOD AUROC of 94.9%, outperforming established methods like Outlier Exposure (OE).
- Synthetic Negatives: Using a Normalizing Flow (Real NVP) to generate synthetic outliers near the inlier manifold boundary proved highly effective. UNO trained with synthetic negatives often outperformed methods using real external negative datasets.
Conclusion
UNO demonstrates that decoupling uncertainty from outlier recognition provides a more robust safety mechanism for deep learning models. By treating "I don't know which class this is" and "I know this is an outlier" as separate problems, the framework achieves superior detection performance without sacrificing in-distribution accuracy. Future work will explore more advanced generative models, such as Diffusion, to further refine synthetic negative generation.
Citation: Delić, A., Grčić, M., & Šegvić, S. (2024). Outlier detection by ensembling uncertainty with negative objectness. arXiv:2402.15374v4 [cs.CV].
Summary updated