Accelerating Video Analytics

MOTIVATION. The advent of inexpensive, high-quality cameras has led to a rapid increase in the volume of generated video data [19, 16]. It is now feasible to automatically analyze these video datasets at scale due to two developments over the last decade. First, researchers have designed complex, computationally-intensive deep learning (DL) models that capture the contents of a given set of video frames (e.g., objects present in a particular frame [11]) [15]. Second, the computational capabilities of hardware accelerators for evaluating these DL models have increased over the last decade (e.g., TPUs) [8]. We anticipate that automated analysis of videos will reduce the labor cost of analyzing video