Outlier detection by ensembling uncertainty with negative objectness


Anja Delić (University of Zagreb), Matej Grcic (Faculty of Electrical Engineering and Computing, University of Zagreb), Siniša Šegvić (UniZg-FER)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Outlier detection is an essential capability of safety-critical visual recognition. Many existing methods deliver good results by encouraging standard closed-set models to produce low-confidence predictions in negative training data. However, that approach conflates prediction uncertainty with recognition of outliers. We disentangle the two factors by revisiting the K+1-way classifier that involves K known classes and one negative class. This setup allows us to formulate a novel outlier score as an ensemble of in-distribution uncertainty and the posterior of the negative class that we term negative objectness. Our UNO score can detect outliers due to either high prediction uncertainty or similarity with negative training data. We showcase the utility of our method in experimental setups with K+1-way image classification and K+2-way dense prediction. In both cases we show that the bias of real negative data can be relaxed by leveraging a jointly trained normalizing flow. Our models outperform the current state-of-the art on standard benchmarks for image-wide and pixel-level outlier detection.

Citation

@inproceedings{Delić_2024_BMVC,
author    = {Anja Delić and Matej Grcic and Siniša Šegvić},
title     = {Outlier detection by ensembling uncertainty with negative objectness},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0779.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection