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 120-3,200× speed-ups (318- 8,500× 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.
In this paper, we address this tension between CNNs’ accuracy and computational expense by employing one of the mainstays of data management system design: query optimization.
Just as conventional database engines automatically adapt their query processing plans to properties of the data , NOSCOPE automatically adapts CNN-based video query processing to the contents of video streams
As a result, NOSCOPE allows efficient and accurate classification and label extraction of data streams—without requiring manual tuning.
In its search for equivalent but efficient CNN filters, NOSCOPE exploits a key property of video data: video content is often highly redundant. NOSCOPE employs two types of filters that exploit several types of locality (i.e., similarities) within a stream. First, video data exhibits temporal locality. Second, video exhibits scene specific spatial locality.
The key challenge in exploiting the above observations is that the optimal configuration of filters is highly data-dependent. Diverse filters will perform very differently on different scenes, offering distinct tradeoffs between speed, selectivity, and accuracy.
To summarize, our contributions are:
1. NOSCOPE, the first data management system that accelerates CNN-based classification queries over video streams at scale.
2. CNN-specific techniques for difference detection across frames and model specialization for a given stream and query, as well as a cost-based optimizer that can automatically identify the best combination of these filters for a given accuracy target.
3. An evaluation of NOSCOPE on fixed-angle binary classification showing up to 3,200× speedups on real-world data.