NTT Communication Science Laboratories and NII in TRECVID 2010 Instance Search Task

Abstract—As a first step to the instance search task, we employed several existing image retrieval methods in combination (local feature matching, region matching, and global feature matching), according to the type and property of query images. The best run in this approach was ranked 23rd of 42 as regards the average precision result. Keywords; PCA-SURF; Bag of Features; Color histogram; Haar-Wavelet; Face Detection; OpenCV I. I NTRODUCTION The instance search task involves locating query topics from a collection of reference videos. The query topics consist of a set of about 5 example frame images, the regions containing the item of interest in the images, the video from which the images were selected, and an indication of the target type taken from the following set of strings: PERSON, CHARACTER, LOCATION, and OBJECT. One collection of reference videos consists of the Sound and Vision data from TRECVID 2009 and each video data is divided into many master shot references. The submitted data comprised 1000 candidates chosen from the master shots for each query topic. The score becomes high when the correct answer is put on the high rank. The similar task with the instance search is the image retrieval from image database [1-3]. This time, we applied several basic existing image retrieval methods according to the type and property of query images. The types of query topic are "PERSON", "CHARACTER", "OBJECT", and "LOCATION". When the query topic type is "PERSON", the region of each query will include the face, and the face information will be important in terms of finding the topic. When the query type is "LOCATION", the region of the query may occupy almost all of the query image, and global information, for example color or frequency information will be useful. On the other hand, the query images are various sizes. Some query images have sufficient size but others have only a few features. So, we adopted a method for selecting the most promising feature for the queries. When there are few features in the query region, we use a method for matching local features, when there are many features in the query region, we use a method for matching the bags of features as a region feature, and when the query region is as large as the query image, we use a method for matching global features. First, we describe our features and their similarity measure in Section II. Section III provides an overview of our system. Section IV reports our submissions and results. Finally, we conclude by some remarks. II.