Research

We predict that applications in AR/VR devices [1] and intelligence devices will lead to the emergence of a new class of image sensors — machine perception CIS (MPCIS). This new class of sensors will produce images and videos optimized primarily for machine vision applications, not human consumption. Unlike human perception CIS, where the ultimate criterion is visual image quality [2], there is no existing criterion to judge MPCIS sensor performance. In this paper, we present a full stack sensor modeling and benchmarking pipeline (from sensors to algorithms) that could serve as the platform for performance evaluation. We illustrate how sensor modeling and benchmarking help us understand complex system trade-offs and dependencies between sensor and algorithm performance, specifically for simultaneous localization and mapping (SLAM).