Contour-HOG: A Stub Feature based Level Set Method for Learning Object Contour

Abstract

In this paper, we address the problem of learning objects by contours. Toward this goal, we propose a novel curve evolution scheme which provides the classifier with more accurate contour representations. We detect edgelet feature to help localize objects in images so that the proposed evolution method can achieve more reliable contours. To capture contours of objects with large variations in pose and appearance, we adopt the similarity measure of HOG feature between the evolving contour and the contour of a class as the evaluation criteria rather than relying on strong shape priors. We encode the joint distribution of the edgelet feature, the HOG feature and the curvature feature of an object in a mixture of Gaussian model. Classification is achieved by computing the posterior of the evolved contour conditioned on the three types of features. Our method is extensively evaluated on the UCF sports dataset, the Caltech 101 dataset, and the INRIA pedestrian dataset. Results show that our method achieves improved performance for the recognition task.

Modeling Spatiotemporal Contextual Dynamics with Sparse-Coded Transfer Learning

“…From this STIR project, we have created a new algorithmic tool set of modeling spatiotemporal contextual dynamics [1][2][3][4][5][6][7][8][9]. For low-level and middle-level visual representation, we proposed a class of Schatten norm based discriminative metrics, locality-constrained low-rank coding, discriminative analysis by multiple principal angles, and clustering based fast low-rank approximation for large scale analysis.…”

“…The evaluation results and outcomes of this research have been published in 8 peer-reviewed conference proceedings along with a best paper ward, and 1 peer-reviewed journal paper. For more details of the research outcomes and accomplishments, please refer to the attached publications [1][2][3][4][5][6][7][8][9]. This report will highlight the details in [3,8].…”