Effective, Stable and Efficient Unsupervised Image Outlier Detection via Distance Ensemble Learning
Abstract: To automatically and efficiently identify whether visual systems involve outliers (anomalies) is an important research topic. Although there has been rapid progress in the efficacy of unsupervised image outlier detection, the instability and complexity of the state-of-the-art (SOTA) methods is still a notable challenge. In this work, we explain the instability problem derived from the mainstream single method-fits-multiple scenarios paradigm, which results in performance fluctuations across different target dataset domains and varying outlier ratios. Therefore, ensembling multiple methods seems necessary. Nevertheless, traditional ensemble learners such as stacking and boosting are less effective without any supervision and are often time-consuming. Such that, we introduce a novel and lightweight distance ensemble learning (DEL) framework featuring self-selection strategies over a series of distance-based methods. Specifically, by exploring a specific property of the high-dimensional space, we propose the normalized Euclidean distance relative to the mean of the target dataset as a reliable baseline. Building upon this baseline method, we enhance it with a conditional bilateral distance metric to achieve stability across diverse dataset domains at low outlier ratios. Furthermore, to address the mean-shift problem encountered by the advanced baseline at high outlier ratios, we integrate it with a high-ratio specific distance transformer, called Shell-Re. This subsequent integration effectively mitigates the advanced baseline's instability across a wide range of outlier ratios. Overall, our approach achieves SOTA results on various challenging benchmarks while offering inference speeds that are orders of magnitude faster.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Philip_K._Chan1
Submission Number: 2557
Loading