Reproduce “Outlier detection by ensembling uncertainty with negative objectness” on OpenResearch

About this paper

Anja Delić, Matej Grcić, Siniša Šegvić

University of Zagreb researchers developed UNO, an outlier detection method that ensembles "negative objectness" and "prediction uncertainty" within an extended K+1-way classification framework. This approach consistently achieves leading or competitive performance on image-wide and pixel-level tasks, such as OpenOOD, Fishyscapes, and SMIYC, often outperforming prior methods and effectively leveraging synthetic negative data.

Key insights

• Prediction uncertainty and negative objectness represent complementary aspects of outliers, with low correlation, and their ensemble in the UNO score leads to a more comprehensive and robust detection capability.
• The disentanglement allows UNO to capture a broader range of outliers, where $s_{\text{Unc}}$ detects ambiguous samples and $s_{\text{NO}}$ identifies samples similar to explicitly trained negatives.
• Synthetic negative samples, generated through a two-step joint training process, can effectively mitigate bias from real negative data and, for some far-OOD cases, even outperform methods using real negatives.

Summary by alphaXiv.