직선 및 Spectral Clustering을 사용한 장소 단위 영상 분류

Abstract

In this paper, we propose a scene grouping algorithm where each resulting group can represent certain concept of a place. We consider each scene as a set of line features and calculate similarities between scenes using a vocabulary tree and TFIDF weighting scheme. Then using the scenes as nodes and their similarities as weighted edges, a similarity graph is constructed. By cutting the sparse edge parts of this graph using spectral clustering, this graph is partitioned into Ɖ groups which correspond to Ɖ places.