Feature Map Matters in Out-of-distribution Detection

Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, Bo Han

Primary Area: societal considerations including fairness, safety, privacy

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: OOD detection

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Detecting and rejecting out-of-distribution (OOD) data can improve the reliability and reduce potential risks of a model (e.g., a neural network) during the deployment phase. Recent post-hoc OOD detection methods usually focus on analyzing hidden features or prediction logits of the model.
However, feature maps of the backbone would also contain important spatial clues for discriminating the OOD data. In this paper, we propose an OOD score function Feature Sim (FS) that can efficiently identify the OOD data by only looking at the feature maps. Furthermore, a novel Threshold Activation (TA) module is proposed to suppress non-critical information in the feature maps and broaden the divergences between foreground and background contexts. We provide a theoretical analysis to help understand our methods. The experimental results show that our methods FS+TA and FS+TA+ASH can achieve state-of-the-art on various benchmarks. More importantly, since our method is based on feature maps instead of hidden features or logits, it can be easily adapted to more scenarios, such as semantic segmentation and object detection.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5135