Test Time Augmentations are Worth One Million Images for Out-of-Distribution Detection

TMLR Paper8580 Authors

Abstract: Out-of-distribution (OOD) detection is commonly improved either by storing large in-distribution (InD) reference sets (e.g., nearest-neighbor methods) or by exposing the model to auxiliary OOD data during training. Both requirements limit deployability at scale. This paper shows that carefully chosen test-time augmentations (TTA) can provide a strong, self-referential signal for OOD detection from a single test input, without any stored InD data and without OOD exposure. We first identify a practical taxonomy that separates mild, feature-preserving InD augmentations (IDAs) from aggressive OOD augmentations (OODAs), and empirically demonstrate that IDAs consistently improve detection while OODAs often degrade it. Building on this insight, we propose a simple plug-and-play detector based on sequential masking: for each test image, we generate a small set of masked views and use the k-th largest embedding similarity to the original image as an “ID-ness” score. With only 25 TTAs per input, our method surpasses competitive baselines on ImageNet that rely on the full 1.2M-image training set as a reference.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Haoliang_Li2

Submission Number: 8580