Scan without a Glance: Towards Content-Free Crowd-Sourced Mobile Video Retrieval System

Mobile videos contain rich information which could be utilized for various applications, like criminal investigation and scene reconstruction. Today's crowd-sourced mobile video retrieval systems are built on video content comparison, and their wide adoption has been hindered by onerous computation of CV algorithms and redundant networking traffic of the video transmission. In this work, we propose to leverage Field of View(FoV) as a content-free descriptor to measure video similarity with little accuracy loss. Based on FoV, our system can filter out unmatched videos before any content analysis and video transmission, which dramatically cuts down the computation and communication cost for crowd-sourced mobile video retrieval. Moreover, we design a video segmentation algorithm and an R-Tree based indexing structure to further reduce the networking traffic for mobile clients and potentiate the efficiency for the cloud server. We implement a prototype system and evaluate it from different aspects. The results show that FoV descriptors are much smaller and significantly faster to extract and match compared to content descriptors, while the FoV based similarity measurement achieves comparable search accuracy with the content-based method. Our evaluation also shows that the proposed retrieval scheme is scalable with data size and can response in less than 100ms when the data set has tens of thousands of video segments, and the networking traffic between the client and the server is negligible.