A Shape Descriptor for Shapes with Boundary Noise and Texture

Abstract

We present a part-based shape descriptor that incorporates both the description of the general shape form of each subpart and its geometric relationship with other connected parts. Associated with each descriptor is a saliency measure that weighs each part's visual significance. By incorporating this saliency measure into the shape matching process, we able to discriminate between shape forms as well as take boundary texture into consideration when computing shape similarity. This paper also describes a multi-resolution pyramidal framework for generating the required gradient vector field and vector field disparity map from which the shape descriptors, in the form of gradient vector field histograms, are derived. Experimental results involving silhouettes images are presented to demonstrate the various characteristics of the proposed shape descriptor, which includes its invariance to similarity transform and its ability to match composite shapes containing boundary noise and texture, limb articulation and occlusion.

A new algorithm for shape matching and pattern recognition using dynamic programming

“…To make the problem even more complex, shape is often corrupted with noise, arbitrary distortion and occlusion. Various shape descriptors [5,6] have been proposed recently.…”

Shape analysis using multiresolution gradient vector field

“…The MGVF skeleton provided the framework by which a 2D shape can be decomposed into skeletal-based parts. An earlier version of the part-based descriptor segmented the skeleton into continuous segments at skeletal junctions [71], [74].…”

“…Earlier versions of this work presented in [71], [73], [74] used this simple part decomposition strategy with some limited success when performing shape classification of a small 25-shapes database used by [18], [65], [195] (see Fig. 4.30).…”

“…This strategy was adopted in [71], [73], [74]. However, it was observed that source node assignment based on maximal disk radial measures did not produce consistent references in shapes such as those shown in Fig 5.33a.…”

“…The second set of MGVF-PBD results using the larger 216-shapes database resulted in recognition rates of (100,100,100,100,100,99,98,97,95,87,74). In other words, the top 5 best matches are always correct when any shape is queried against the remaining 215 shapes in the database.…”