TASM: A Tile-Based Storage Manager for Video Analytics
share
The amount of video data being produced is rapidly growing. At the same time, advances in machine learning and computer vision have enabled applications to query over the contents of videos. For example, an ornithology application may retrieve birds of various species from a nature video. However, modern video data management systems store videos as a single encoded file, which does not provide opportunities to optimize queries for spatial subsets of videos. We propose utilizing a feature in modern video codecs called "tiles" to enable spatial random access into encoded videos. We present the design of TASM, a tile-based storage manager, and describe techniques it uses to optimize the physical layout of videos for various query workloads. We demonstrate how TASM can significantly improve the performance of queries over videos when the workload is known, as well as how it can incrementally adapt the physical layout of videos to improve performance even when the workload is not known. Layouts picked by TASM speed up individual queries by an average of 51 to 94
READ FULL TEXT