Local phase quantization descriptor for improving shape retrieval/classification


Shape classification is a field of study with applications ranging from object classification to leaf recognition. In this paper we present an approach based on a matrix descriptor, the local phase quantization, for improving the performance of such widely used shape descriptors as the inner distance shape context (ID), shaper context (SC), and height functions (HF). The basic idea of our novel approach is to transform the shape descriptor obtained by ID/SC/HF into a matrix so that matrix descriptors can be extracted. These matrix descriptors are then compared with the Jeffrey distance and combined with standard ID/SC/HF shape similarity. Since it has recently been shown that ID/SC/HF shape similarities can be coupled with learning context-sensitive shape similarity using graph transduction (LGT) for improving the results, we have also coupled our approach with LGT. Our proposed approach is tested on a wide variety of shape databases including MPEG7 CE-Shape-1, Kimia silhouettes, a leaf dataset, a tools dataset, a myths figures dataset, and our new human dancer dataset. The experimental results demonstrate that the proposed approach yields significant improvements over baseline shape matching algorithms. All Matlab code used in the proposed paper is available at bias.csr.unibo.it\nanni\Shape.rar.

Keywords: Shape classification; local phase quantization; inner distance shape context; shape context; height functions; texture descriptors.

[full paper]