Learning Accurate and Human-Like Driving using Semantic Maps and Attention

This paper investigates how end-to-end driving models can be improved to\ndrive more accurately and human-like. To tackle the first issue we exploit\nsemantic and visual maps from HERE Technologies and augment the existing\nDrive360 dataset with such. The maps are used in an attention mechanism that\npromotes segmentation confidence masks, thus focusing the network on semantic\nclasses in the image that are important for the current driving situation.\nHuman-like driving is achieved using adversarial learning, by not only\nminimizing the imitation loss with respect to the human driver but by further\ndefining a discriminator, that forces the driving model to produce action\nsequences that are human-like. Our models are trained and evaluated on the\nDrive360 + HERE dataset, which features 60 hours and 3000 km of real-world\ndriving data. Extensive experiments show that our driving models are more\naccurate and behave more human-like than previous methods.\n