preprints.org > engineering > automotive engineering > doi: 10.20944/preprints202408.2045.v1 (registering DOI)

Preprint Article Version 1 This version is not peer-reviewed

Version 1 : Received: 23 August 2024 / Approved: 28 August 2024 / Online: 28 August 2024 (10:38:00 CEST)

Recent research has drawn attention to the ambiguity surrounding the definition and learnability of Out Of Distribution (OOD) recognition. The term "Out of Model Scope" (OMS) detection provides a clearer perspective, although the original problem remains unsolved. The ability to detect OMS inputs is particularly beneficial in safety-critical applications such as autonomous driving or medicine. By detecting OMS situations, we not only enhance the system's robustness but also prevent it from operating in unknown and unsafe scenarios. In this paper, we introduce a novel approach for OMS detection that combines three sources of information: 1) the original input, 2) the latent feature space extracted by an encoder, and 3) the synthetic data generated from the encoder's latent feature space. We demonstrate the effectiveness of combining original images and synthetically generated images against adversary attacks in the computer vision domain. We achieve results comparable to other state-of-the-art methods and demonstrate that any Encoder can be integrated into our pipeline in a plug-and-train fashion. Our experiments further evaluate which combination of the Encoder's features works best in order to discover OMS samples and highlight the importance of a compact feature space for the training of the Generator.

Out of Distribution (OOD) Recognition; Safety-Critical Applications; Generator as Distribution Approximator; Robustness of Computer Vision Systems

Engineering, Automotive Engineering

* All users must log in before leaving a comment