SysML’18, February 15–16, 2018, Stanford, CA.
Christopher Canel, Thomas Kim, Giulio Zhou, Conglong Li, Hyeontaek Lim, David G. Andersen, Michael Kaminsky†, Subramanya R. Dulloor†
Carnegie Mellon University
† Intel Labs
As video camera deployments proliferate in the smart cities of the future , software systems are faced with the increasing challenge of determining which segments of data are relevant. For resource-constrained edge nodes, limited network bandwidth back to the datacenter prevents sending entire video streams.
This paper presents a new application-independent interesting frame (IF) detection algorithm for identifying relevant frames in streaming video. We envision this IF detector as a preprocessing step in a larger video analytics pipeline where the expensive computation occurs later (similar to the way Bloom filters can guard expensive data structures). Given a target frame rate (or, equivalently, a target bandwidth), the algorithm decides which frames are the most generally interesting and therefore should be processed by downstream applications or forwarded to the datacenter.We decide how “interesting” a frame is based on its semantic difference from other frames. The IF detection algorithm uses a hierarchy of filters to trade off between end-to-end latency and aggressive decimation. The algorithm strives to maximize the semantic diversity of the selected frames. Compared to simply choosing frames at a fixed interval, the IF detector better handles bursty events in the stream.
FULL PAPER: pdf