We present content-based video retrieval using image queries that show a specific person/object/place. This type of searching is called “instance search” and it has been actively discussed in the TREC Video Retrieval Evaluation (TRECVID) community since 2010. Our approach is based on a probabilistic information retrieval model, and it generates search results using bags of local visual features extracted from image queries and videos in a database. These features correspond to keypoints and by using the novel keypoint-weighting method, our instance search results currently record the world’s highest level in search accuracy for a series of the TRECVID dataset. This research contributes to work on robust and smart media search technology that enables responding to the complex search intentions of today’s users.
Poster
Please click the thumbnail image to open the full-size PDF file.