Image-Based Video Retrieval Using Deep Feature

In this paper, we focus on retrieving video by an image querying. Current approaches involve extracting hand- craft features from each key-frame of videos, which is memory cost. We propose to use deep feature deriving from deep neural network to tackle this issue. Specifically, deep feature is employed to detect shots consisting of similar key-frames and represent them by different aggregation strategies, which can avoid saving redundant key-frames of videos. In addition, to discount the contribution of background, we propose a two-way localization approach, which searches the best matched regions between query and video key-frames. Then, the updated similarity built upon the best matched regions is utilized to re-rank initial retrieval results for further refinement. Experimental results over the public CNN2h dataset demonstrate the effectiveness of the proposed approach.