Probabilistic predicates to accelerate inference queries - Microsoft Research

Imagine having to narrow down a video of passing vehicles captured on a traffic camera to red SUVs that are exceeding the speed limit. Such queries over images, videos and text are an increasingly common use-case for data analytics platforms. A query to identify in-a-hurry red SUVs has to execute several machine learning modules on the video, for example, to detect parts of a video frame that contain vehicles, to classify the vehicles by type (SUV, truck, and so on), to track vehicles across frames and to estimate the speed of every moving vehicle. Eventually, a predicate picks the frames that have red SUVs traveling above the speed-limit. Unfortunately, such queries are slow today because the machine learning modules can be computationally intensive.