[DB Seminar] Spring 2017: Alex Poms

A growing number of visual computing applications depend on the analysis of large video collections. The challenge is that scaling applications to operate on these datasets requires highly efficient systems for pixel data access and parallel processing. Few programmers have the capability to operate efficiently at these scales, limiting the field’s ability to explore new applications that analyze large video data sets.

Inspired by the impact of systems such as analytics databases and Spark, we are developing Scanner, a platform for productive and efficient video analysis at scale. Scanner organizes video collections as relations (tables) in a data store, and facilitates efficient delivery of video frame data to pixel processing pipelines that operate on these collections. Scanner’s design emphasizes high efficiency, utilizing heterogeneous throughput computing hardware, such as GPUs and media processing ASICs, for high-throughput pixel processing.

We demonstrate preliminary results suggesting the utility of Scanner in several application scenarios, including cinematography studies on 500 feature-length films, markerless 3D human pose reconstruction from video, and rapidly labeling a dataset of 100,000 videos. We will also give a live demonstration of the system on a cluster of 8-16 GPUs.

Bio:
Alex Poms is a second year Ph.D. student at CMU who is building systems for large-scale visual data analysis. His primary research focus is in combining ideas from HPC, distributed systems, and databases to organize high-throughput visual computing. A long long time ago, he worked on multi-GPU, multi-node solvers for convex optimization with Prof. John D. Owens & Prof. Stephen Boyd. His Ph.D. advisor is Prof. Kayvon Fatahalian.