Training Decisions: Ground-Truthing the Interesting

This chapter examines two training datasets—the KITTI Vision Benchmark Suite and Waymo Open Dataset—constituting key milestones in the ‘ground truthing’ of autonomous vehicles. Launched only seven years apart, each dataset is nonetheless markedly different, constitutive of very different moments in the development of autonomous vehicles—what I have called here the benchmark and incremental phases of autonomous driving. Often taken-for-granted, training datasets are integral components in the development of autonomous vehicles, critical to the building of ML models that underpin their executive functions. Only ever as useful as the classification and annotation work performed on them, autonomous vehicle training datasets—such as the KITTI Vision Benchmark Suite and Waymo Open Dataset—demonstrate great variety in composition, technical set-up, source data, and volume, amongst many other things. The KITTI Vision Benchmark Suite, launched in 2012, offered the first real-world benchmarks for the comparison of machine vision systems used in autonomous driving settings. An emerging interest in ‘interesting-ness’, typified by Waymo engineers’ repeated emphasis on scenario diversity within such datasets, similarly defines the incremental era of autonomous driving, committed to refining extant machine learning and machine vision techniques.