Sensing Decisions: Perceiving, Classifying, Finessing

This chapter explores the ‘sensor strategies’ being devised to optimize object-recognition processes, critical to gifting autonomous vehicles their perceptive qualities. Although such sensor work is already necessary for the capture of ML training data in the first instance, novel methods and approaches are likewise required in order to improve what researchers call the ‘online’ (i.e. real-time) capabilities of their object-recognition systems. Drawing on work on the ‘operational’ nature of digital images and data, the chapter considers how machine vision researchers, such as those at Argo AI, are engaged in various efforts to rationalize the object-recognition process. Offering practical examples of the need for ‘interoperability’ between different stages in the autonomous vehicle decision-making process, such work demonstrates the need for interoperability on different technical, epistemological, and organizational levels. Practical examples encountered during autonomous driving and machine vision workshops include innovative, interstitial techniques to upgrade lidar to a fully 3D sensing format and ‘dynamic scheduling’ techniques to balance quick and accurate image understanding. In both cases the chapter understands such work as integral to ‘finessing’ the interoperability of autonomous vehicle systems, ensuring ingested sensor data is prepared for subsequent stages in the decision-making pipeline.