Boggart - Towards General-Purpose Acceleration of Retrospective Video Analytics

Explore a conference talk that introduces Boggart, a system designed to accelerate retrospective video analytics while maintaining generality, accuracy, and efficiency. Delve into the challenges faced by commercial platforms in supporting custom queries and convolutional neural networks (CNNs) for various applications. Discover how Boggart employs traditional computer vision algorithms to create comprehensive indices prior to query issuance, and learn about its innovative techniques for characterizing index imprecision and selectively running CNNs to bound accuracy drops. Gain insights into Boggart's performance, which matches or exceeds prior model-specific approaches while offering improved generality at low cost.