Towards Explainable Semantic Segmentation for Autonomous Driving Systems by Multi-Scale Variational Attention

Abstract

Explainable autonomous driving systems (EADS) are emerging recently as a combination of explainable artificial intelligence (XAI) and vehicular automation (VA). EADS explains events, ambient environments, and engine operations of an autonomous driving vehicular, and it also delivers explainable results in an orderly manner. Explainable semantic segmentation (ESS) plays an essential role in building EADS, where it offers visual attention that helps the drivers to be aware of the ambient objects irrespective if they are roads, pedestrians, animals, or other objects. In this paper, we propose the first ESS model for EADS based on the variational autoencoder (VAE), and it uses the multiscale second-order derivatives between the latent space and the encoder layers to capture the curvatures of the neurons’ responses. Our model is termed as Mgrad2 VAE and is bench-marked on the SYNTHIA and A2D2 datasets, where it outperforms the recent models in terms of image segmentation metrics.