high performance computing on graphics processing units: hgpu.org

Video is one of the fastest-growing sources of data and is rich with interesting semantic information. Furthermore, recent advances in computer vision, in the form of deep convolutional neural networks (CNNs), have made it possible to query this semantic information with near-human accuracy (in the form of image tagging). However, performing inference with state-of-the-art CNNs is computationally expensive: analyzing videos in real time (at 30 frames/sec) requires a $1200 GPU per video stream, posing a serious computational barrier to CNN adoption in large-scale video data management systems. In response, we present NOSCOPE, a system that uses cost-based optimization to assemble a specialized video processing pipeline for each input video stream, greatly accelerating subsequent CNNbased queries on the video. As NOSCOPE observes a video, it trains two types of pipeline components (which we call filters) to exploit the locality in the video stream: difference detectors that exploit temporal locality between frames, and specialized models that are tailored to a specific scene and query (i.e., exploit environmental and query-specific locality). We show that the optimal set of filters and their parameters depends significantly on the video stream and query in question, so NOSCOPE introduces an efficient cost-based optimizer for this problem to select them. With this approach, our NOSCOPE prototype achieves up to 120-3,200x speed-ups (318- 8,500x real-time) on binary classification tasks over real-world webcam and surveillance video while maintaining accuracy within 1-5% of a state-of-the-art CNN.