Object Categorization Utilizing a Codebook Containing Contextual Information of Visual Words

情報処理学会

説明

In object categorization, bag of visual words is a promising approach. However, in this framework how to obtain discriminative codebook is still an open issue. Since contextual information can be used to reduce ambiguity in object recognition, in this report we propose to build a codebook which takes contextual information of visual words into consideration. Utilizing a codebook in which both the appearance of visual words and their contextual information are contained would help to improve image representation. We first detect interest points in images employing Harris-Laplacian detector, then from each detected point we extract patches of different scales, which are described using SIFT descriptor. After that, based on these extracted patches we build a hierarchical codebook in which visual words in different levels are related, and higher level visual words contain contextual information of lower level visual words. Through this codebook, image representations which are more discriminative and robust could be created. We compared our method with two baseline approaches, and results indicated the effectiveness of our proposed method.

収録刊行物

詳細情報 詳細情報について

問題の指摘

ページトップへ