Date: Thursday, May 22
Start Time: 4:15 pm
End Time: 4:45 pm
Electronic products for virtual and augmented reality, home robots and cars deploy multiple cameras for computer vision and AI use cases. These cameras are used for detection, tracking, recognition, SLAM, biometric authentication and a variety of other use cases, all enabled via traditional computer vision algorithms or neural networks. Camera frames produced for these algorithms do not need sophisticated image signal processing (such as that performed in ISP hardware) if the neural network models were trained on data collected without use of such image processing. In addition, some camera functionality supported for photography applications (such as extensive metadata and the ability to change camera settings on a per-frame basis) may not be needed for AI use cases. These differing requirements create an opportunity to optimize the camera stack for better performance, reducing CPU load. In this talk, we will share optimization strategies to create a lightweight AI camera stack that can be executed on low-power devices for AI use cases.
