稀疏雜湊學習法與語意標註探勘應用於行動裝置上之大規模影片搜尋

Abstract

Retrieving relevant videos from a large corpus is a long-standing research problem, and doing so on mobile devices brings additional technical challenges. This paper addresses two key issues for mobile applications on user-generated videos. The first is the lack of good relevance measurement, due to the unconstrained nature of online videos. The second is the strict requirement on efficiency, due to the limited resource on mobile device, stringent bandwidth and delay requirement between the device and the video server. We propose two novel approaches for each problem. We carry out Pseudo Label Mining based on Explicit Semantic Analysis to generate high quality similar video pairs. This method connects the video metadata with Wikipedia semantics, and alleviates the need for expensive annotated data. In addition, we propose a novel sparse projection method to address the efficiency challenge. It learns a discriminative compact representation that drastically reduces transmission cost. With less than 10% non-zero element in projection matrix, it also reduces computational and storage cost. The experimental results on 100k videos show that our proposed algorithm is competitive in the MAP performance to the state-of-the-art semi-supervised hashing method which is not applicable on mobile platforms. The average query time on 100k videos consumes only 0.592 seconds.