Reliability Evaluation of Visualization Performance of Convolutional Neural Network Models for Automated Driving

Abstract

As deep learning methods in image recognition have achieved excellent performance, researchers have begun to apply CNNs(convolutional neural networks) to automated driving. However, the explainability for the decision making of automated driving is highly desired. In order to trust the model in automated driving, visualization methods have become important for understanding the internal calculation process of CNNs. Therefore, in a previous study, we proposed a method to evaluate the visualization performance of CNN models by using a mathematical model instead of a human driver to generate a dataset that can determine the ground-truth point in images. However, the reliability of the proposed method for validating the visualization performance was not provided. Therefore, in this paper, we verify the proposed method through two experiments to demonstrate the task-dependent performance and visualization performance during training. The reliability of the visualization performance has been demonstrated through experimental results. Therefore, we proposed an evaluation method for visualization performance in automated driving systems.