Weighted Fusion of Shape Descriptor for Robust Shape Classification


Shape classification in computer vision is a vibrant field of study with wide ranging applications involving object classification, motion tracking, image segmentation, and image retrieval. In this work we propose an ensemble of approaches harnessing the power of many shape descriptors (inner distance shape context, shape context, height functions, etc.) that are transformed into a matrix from which a set of texture descriptors are extracted and compared using the Jeffrey distance. The experimental results confirm that our new ensemble of texture descriptors clearly improve previous texture-based ensembles for shape classification. Our ensemble is validated and compared with the state-of-the-art on several benchmark datasets, representing different shape classification tasks (MPEG7 CE-Shape-1, Kimia silhouettes, Tari dataset, a leaf dataset, and an animal shape dataset). The parameters of each method used in our ensemble and the weights of the weighted fusion by sum rule are kept the same across each of these datasets, thereby demonstrating that our proposed ensemble is a general-purpose shape classification system. Moreover, the texture descriptors based approach is also coupled with the Bag of Contour Fragments method permitting a boost of the performance obtained by the standard approach. The MATLAB code of our proposed system will be available online at https://www.dei.unipd.it/node/2357.

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

[full paper]