BlazeIt: Fast Exploratory Video Queries using Neural Networks
share
As video volumes grow, analysts have increasingly turned to deep learning to process visual data. While these deep networks deliver impressive levels of accuracy, they execute as much as 10x slower than real time (3 fps) on a 8,000 GPU, which is infeasible at scale. In addition, deploying these methods requires writing complex, imperative code with many low-level libraries (e.g., OpenCV, MXNet), an often ad-hoc and time-consuming process that ignores opportunities for cross-operator optimization. To address the computational and usability challenges of video analytics at scale, we introduce BLAZEIT, a system that optimizes queries over video for spatiotemporal information of objects. BLAZEIT accepts queries via FRAMEQL, a declarative language for exploratory video analytics, that enables video-specific query optimization. We propose new query optimization techniques uniquely suited to video analytics that are not supported by prior work. First, we adapt control variates to video analytics and provide advances in specialization for aggregation queries. Second, we adapt importance-sampling using specialized NNs for cardinality-limited video search (i.e. scrubbing queries). Third, we show how to infer new classes of filters for content-based selection. By combining these optimizations, BLAZEIT can deliver over three order of magnitude speedups over the recent literature on video processing.
READ FULL TEXT
