Hidden-Layer Monitoring for Out-of-Distribution Localization in Image Segmentation

Jan Křetínský, Sabine Rieder, Gesina Schwalbe, Youssef Shoeb

Abstract: The superior performance of neural networks (NNs) in safety-critical situations like automated driving (AD) vision is challenged by so-called out-of-distribution (OoD) examples: These are samples that are improbable according to the training data’s distribution, such as novel object classes. In such unusual cases, NNs are prone to produce erroneous predictions with high confidence. Therefore, real-time capable localization of OoD areas in input images is needed, enabling appropriate caution if OoD locations conflict with the AD trajectory. A promising direction is computationally efficient hidden-layer “distribution-based” OoD monitoring methods. They model the activation values of neurons in a given hidden layer of the NN (so-called latent features) using probability distributions. During runtime, they then flag images yielding low probability as OoD. These methods have been successfully applied to classification, but neither to OoD localization nor object segmentation NNs. This paper investigates how far these monitoring techniques can be adapted to OoD localization and performs an extensive case study with several monitoring techniques. Additionally, we examine potential influence factors like NN architecture and training data. Our results demonstrate that this is a promising direction for efficient OoD localization.

External IDs:doi:10.1007/978-3-032-05435-7_16