Design and Calibration of Multi-camera Systems for 3D Computer Vision: Lessons Learnt from Two Case Studies

摘要

This paper examines how the design of imaging hardware for multi-view 3D reconstruction affects the performance and complexity of the computer vision system as a whole. We examine two such systems: a grape vine pruning robot (a 4.5 year/20 man-year project), and a breast cancer screening device (a 10 year/25 man-year project). In both cases, mistakes in the initial imaging hardware design greatly increased the overall development time and cost by making the computer vision unnecessarily challenging, and by requiring the hardware to be redesigned and rebuilt. In this paper we analyse the mistakes made, and the successes experienced on subsequent hardware iterations. We summarise the lessons learned about platform design, camera setup, lighting, and calibration, so that this knowledge can help subsequent projects to succeed.