Finding relevant videos in big data environments - how to utilize graph processing systems for video retrieval

Abstract

The fast growing amount of videos in the web arises new challenges. The first is to find relevant videos for specific queries. This can be addressed by Content Based Video Retrieval (CBVR), in which the video data is used to do retrieval. A second challenge is to perform such CBVR with big amounts of data. In this work both challenges are targeted by using a distributed Big Graph Processing System for CBVR. A graph framework for CBVR is built with Apache Giraph. The system is generic in regard of the used feature set. A similarity graph is built with the chosen features. The graph system provides a insert operation for adding new videos and a query operation for retrieval. The query uses a fast fuzzy search for seeds of a personalized Pagerank, which uses the locality of the similarity graph for improving the fuzzy search. The graph system is tested with SIFT features for object recognition and matching. In the evaluation the Stanford I2V is used.