Discovering Motifs in a Database of Shapes under Rotation Invariant Dynamic Time Warping

Abstract

Finding similar two-dimensional shapes is an important problem with application in various domains. Shape motifs are approximately repeated shapes within image collections. Two-dimensional shapes can be converted to time series and the problem of discovering motifs in shape time series can be solved by exploiting some appropriate techniques from time series motif discovery. However, so far there exists one method proposed by Xi et al. which uses Euclidean distance to deal with this problem. As for shape data, Dynamic Time Warping (DTW) distance is more suitable and brings out more accurate results than Euclidean distance in similarity matching. Nevertheless, there are two difficulties with DTW distance in image matching: DTW distance incurs high computational complexity and image matching requires that the distance measure must be invariant to rotation. In this work, we propose a fast method for discovering shape motifs in a database of shapes using rotation invariant DTW distance. Our method is not based on Random Projection algorithm but employs a new approach for motif-discovery: clustering-based. Experimental results showed that our proposed method for shape motif discovery under DTW performs very efficiently on large datasets of shapes while brings out high accuracy.