Autonomous vehicles requires good learning models which, in turn, require a large amount of real-world sensor training data. Unfortunately, the staggering volume of data produced by in-vehicle sensors, especially the cameras, make both local storage and transmission of this data to the cloud for training prohibitively expensive. In this work, we explore techniques for reducing video frames in a way that the quality of training for autonomous vehicles is minimally affected. We particularly focus on utility aware data reduction schemes where the potential contribution of a video frame to enhancing the quality of learning (or utility) is explicitly considered during data reduction. Since actual utility of a video frame cannot be computed online, we use surrogate utility metrics to decide what video frames to keep for training and which ones to discard. Our results show that utility-aware data reduction schemes can reduce the amount of camera data required for training by as much as 16× compared to random sampling for the same quality of learning (in terms of IoU).