Live Video Analytics at Scale with Approximation and Delay-Tolerance

Explore a conference talk on VideoStorm, a groundbreaking video analytics system designed to process thousands of queries on live video streams across large clusters. Dive into the challenges of resource management in vision processing, focusing on resource-quality tradeoffs and varying quality and lag goals. Learn about VideoStorm's innovative offline profiler and online scheduler, which optimize performance by allocating resources based on query requirements rather than traditional fair sharing methods. Discover how this system achieves up to 80% improvement in quality and 7x better lag for real-world queries, demonstrated through deployment on an Azure cluster of 101 machines. Gain insights into the potential applications for security and smart city scenarios, where millions of cameras generate vast amounts of live video data requiring efficient, real-time analysis.