With the coming of mobile age, huge amount of multimedia files from the smartphone users are distributed on the Internet. We are now in urgent need of an efficient distributed platform to process and analyze those multimedia data. However, existing approaches typically suffer from compatibility issues with legacy multimedia understanding implementations; memory management issues due to the large file size of multimedia data; and limited platform data format support. In this talk, we present a platform level design and implementation to facilitate large scale distributed analysis and understanding of multimedia files, tackling difficulties mentioned above through (1) binary data pipe based execution (2) all data processing contained inside distributed platform (3) streaming based implementation, and (4) flexible I/O types to support various application scenarios. We will demonstrate how this framework is used in our image monetization product to speed up our model training pipeline and improve our CTR prediction.
Quan Wang is currently a software architect of the big data infrastructure team at Baidu USDC working on distributed feature extraction and model training for multimedia data based on Spark. Prior to that, he worked on virtualization projects for high performance networking in Ericsson. Before joining Ericsson, Quan obtained his Ph.D. degree from USC focused on Machine Learning and Pattern Recognition, and before that he studied Mathematics and Mechanics in China.