マルチモーダルなコンテンツ解析に基づく撮影地点の推定 地名の固有表現抽出と画像照合に基づく地点探索の併用

抄録

A combined geo-search method to estimate characteristic location s appearing in given video content is proposed. The combined approach is composed of two stages: a geo-coding stage based on named entity extraction of each place showing in the closed captions, and a scenery image matching stage using Google Street View panorama images of those places. In the second stage, cosine similarities of GIST descriptor s contributes to reducing estimation errors compared to the reciprocal mean di stance of matched pairs of akaze descriptors and color histogram correlation.