Baseline approach for instance search task: local region-based face matching and regional combination of local features (パターン認識・メディア理解)

概要

論文の詳細を見る
The instance search task is introduced as a pilot task in TRECVID 2010. This task formulates the need of searching by visual examples that is crucial in many video search engines. The system is provided by images and regions of the entity of interest and is required to return a ranked list of 1,000 shots that are likely to contain an instance of the entity. Four targeted entity types, PERSON, CHARACTER, OBJECT and LOCATION, present various query types that users usually perform searching. In this paper, we present an approach that can be considered as the baseline for the instance search task. Our approach uses different feature types for different query types. Specifically, we tested two feature types that have been using widely in face recognition applications for PERSON and CHARACTER queries: pixel intensity extracted around facial feature points and local binary patterns extracted from sub-regions of a face. As for the remaining types of queries such as LOCATION, OBJECT, we tested global features (color), local features (SIFT), and the combination of global and local features. Our system obtained good performance on query types such as PERSON and CHARACTER (ranked 1st), and reasonable performance on the remaining query types such as LOCATION and OBJECT.
社団法人電子情報通信学会の論文
2011-02-10