Finding Event Videos via Image Search Engine

Searching desirable events in uncontrolled videos isa challenging task. Current researches mainly focus on obtaining concepts from numerous labeled videos. But it is time consumingand labor expensive to collect a large amount of required labeled videos to model events under various circumstances. To alleviate the labeling process, we propose to learn models for videos by leveraging abundant Web images which contains a rich source of information with many events taken under various conditions and roughly annotated. However, knowledge from the Web is noisy and diverse, brute force knowledge transfer may hurt the retrieval performance. To address such negative transfer problem, we propose a novel Joint Group Weighting Learning (JGWL) framework to leverage different but related groups of knowledge (source domain) queried from the Web image searching engine to real-world videos (target domain). Under this framework, weights of different groups are learned in a joint optimization framework, and each weight represents how contributive the corresponding image group is to the knowledge transferred to the videos. Moreover, to deal with the feature distribution mismatching between video feature space and image feature space, we build a common feature subspace to bridge these two heterogeneous feature spaces in an unsupervised manner. Experimental results on two challenging video datasets demonstrate that it is effective to use grouped knowledge gained from Web images for video retrieval.