Scalable Mobile Video Retrieval with Sparse Projection Learning and Pseudo Label Mining

Retrieving relevant videos from a large corpus on mobile devices is a vital challenge. This article addresses two key issues for mobile search on user-generated videos. The first is the lack of good relevance measurement for learning semantically rich representations, due to the unconstrained nature of online videos. The second is the limited resources on mobile devices, stringent bandwidth, and delay requirement between the device and video server. The authors propose a knowledge-embedded sparse projection learning approach. To alleviate the need for expensive annotation in hash learning, they investigate varying approaches for pseudo label mining, where explicit semantic analysis leverages Wikipedia. In addition, they propose a novel sparse projection method to address the efficiency challenge by learning a discriminative compact representation that drastically reduces transmission costs. With less than 10 percent nonzero elements in the projection matrix, it also reduces computational and storage costs. The experimental results on 100,000 videos show that the proposed algorithm yields performance competitive with the prior state-of-the-art hashing methods, which are not applicable for mobiles and solely rely on costly manual annotations. The average query time for 100,000 videos was only 0.592 seconds.