Ensemble of shape descriptors for shape retrieval and classification

[abstract]

Shape classification has long been a field of study in computer vision. In this work we propose an ensemble of approaches using the weighted sum rule that is based on a set of widely used shape descriptors (inner distance shape context, shape context, and height functions). Features are obtained by transforming these shape descriptors into a matrix from which a set of texture descriptors are extracted. The different descriptors are then compared using the Jeffrey distance. We validate our ensemble on seven widely used datasets (MPEG7 CE-Shape-1, Kimia silhouettes, Tari dataset, a leaf dataset, a tools dataset, a myths figures dataset, and motif pottery dataset), where the parameters of each method and the weights of the weighted fusion are kept the same across all seven datasets, thereby producing a general-purpose shape classification system. Our experimental results demonstrate that our new generalized approach offers significant improvements over baseline shape matching algorithms.

Keywords: shape classification; ensemble; weighted sum rule; Jeffrey distance; texture descriptors.