This study tackles the problem of robust novelty detection, specifically aiming to detect semantic novelties while ignoring irrelevant stylistic changes. The authors propose a method that starts with a pretrained embedding and a multi-environment setup, ranking features based on their environment-focus. The method computes a per-feature score based on feature distribution distances between environments. The study introduces a synthetic benchmark for robust novelty detection, COCOShift, and validates the results on the DomainNet dataset. The method reportedly improves overall performance by up to 6% by removing spurious correlations.
Publication date: 21 Sep 2023
Project Page: https://arxiv.org/abs/2309.12301v1
Paper: https://arxiv.org/pdf/2309.12301
