Object detection and classification using shape feature

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] We develop a set of methods to represent and detect shapes in images. We first develop new shape descriptors that are robust to deformation while being able to capture part details. In our framework, the shape descriptor is generated by 1) using running angle to transforming a shape into a 2-D description image in the position and scale space; 2) performing circular wavelet-like sub-band decomposition of this 2-D description image based on its periodic convolution with orthogonal kernel functions. The shapes are classified with linear SVM. Our performance evaluations on several public datasets demonstrate that the proposed method significantly outperforms state-of-the-art methods. We then study the problem of detecting deformable objects from cluttered images given a single object sketch as model. To address this challenge, we develop local shape descriptors and additive similarity metric function which can be computed locally while preserving the capability of matching deformable shapes globally. To effectively detect objects with large deformation, we augment the metric function with local motion search, model the relationship between different shape parts using multiple concurrent dynamic programming shape parsers, and finalize the detection result using Hough voting. Our experimental results show that the proposed method outperforms the state-of-the-art shape-based object detection algorithms on the benchmark datasets in terms of average precision.