Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail

Yina He,Lei Peng,Yongcun Zhang,Juanjuan Weng,Zhiming Luo,Shaozi Li

2024-08-24

Abstract:Current out-of-distribution (OOD) detection methods typically assume balanced in-distribution (ID) data, while most real-world data follow a long-tailed distribution. Previous approaches to long-tailed OOD detection often involve balancing the ID data by reducing the semantics of head classes. However, this reduction can severely affect the classification accuracy of ID data. The main challenge of this task lies in the severe lack of features for tail classes, leading to confusion with OOD data. To tackle this issue, we introduce a novel Prioritizing Attention to Tail (PATT) method using augmentation instead of reduction. Our main intuition involves using a mixture of von Mises-Fisher (vMF) distributions to model the ID data and a temperature scaling module to boost the confidence of ID data. This enables us to generate infinite contrastive pairs, implicitly enhancing the semantics of ID classes while promoting differentiation between ID and OOD data. To further strengthen the detection of OOD data without compromising the classification performance of ID data, we propose feature calibration during the inference phase. By extracting an attention weight from the training set that prioritizes the tail classes and reduces the confidence in OOD data, we improve the OOD detection capability. Extensive experiments verified that our method outperforms the current state-of-the-art methods on various benchmarks.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper aims to address the issue of Out-of-Distribution (OOD) detection in long-tailed distribution data. Most current OOD detection methods assume that the training dataset is balanced, whereas in real-world applications, data often follows a long-tailed distribution. Previous long-tailed OOD detection methods typically balance the data by reducing the semantic information of head classes, but this approach severely affects the classification accuracy of normal classes.

To solve this problem, the paper proposes a new method called "Prioritizing Attention to Tail" (PATT). The core of this method includes:
1. **Implicit Semantic Augmentation Contrastive Learning (ISAC)**: Utilizing von Mises-Fisher (vMF) distribution to generate an infinite number of contrastive samples, thereby balancing the long-tailed distribution data and improving the distinction between normal and abnormal classes.
2. **Temperature Scaling-Based Logit Adjustment (TLA)**: Ensuring high confidence for normal classes to enhance OOD detection capability.
3. **Post-Hoc Feature Calibration (FC)**: Extracting attention weights during the inference phase to prioritize the features of tail classes and reduce the confidence of abnormal classes.

Experimental results show that the PATT method significantly outperforms existing methods on multiple benchmark datasets (such as CIFAR10-LT, CIFAR100-LT, and ImageNet-LT), with notable improvements in AUROC and normal class classification accuracy. This indicates that the PATT method has a stronger comprehensive semantic representation capability when dealing with long-tailed distribution data, making it more suitable for real-world application scenarios.