More Persuasive Explanation Method for End-to-End Driving Models

With the rapid development of autonomous driving technology, a variety of high-performance end-to-end driving models (E2EDMs) are being proposed. In order to understand the computational methods of E2EDMs, pixel-level explanations methods are used to obtain the explanations of the E2EDMs. However, little attention has been paid to the excellence of the explanations of E2EDMs. Therefore, in order to build trustworthy E2EDMs, we focus on improving the persuasibility of the explanations of E2EDMs. We propose an object-level explanation method (main approach) for E2EDMs, which masks the objects in the image and then treats the change in the prediction result as the importance of the objects, then we explain the E2EDM by the importance of each object. To further validate the effectiveness of object-level explanations, we propose another approach (validation approach), which trains E2EDMs with object information as input and generates the importance of objects using general explanation methods. Both approaches generate object-level explanations, in order to compare these object-level explanations with traditional pixel-level explanations, we propose experimental methods to measure the persuasibility of explanations of E2EDMs through a subjective and objective method. The subjective method evaluates persuasibility based on the extent to which participants think the importance of features indicated by the explanations is correct. The objective method evaluates the persuasibility based on the human annotation similarity between provided with only the important part of images and provided with the complete images. The experimental results show that the object-level explanations are more persuasive than the traditional pixel-level explanations.