Affiliations:
Department of Computer Science, Utah State University, Logan, UT, USA
Correspondence:
[*]
Corresponding author. Liang Peng, Department of Computer Science, Utah State University, Logan, UT, USA. Tel.: +(785)236 1972; Fax: +(435)797 3265; E-mail: [email protected].
Abstract: Generating class-agnostic object proposals followed by classification has recently become a common paradigm for object detection. Current state-of-the-art approaches typically generate generic objects, which serve as candidates for object classification. Since these object proposals are generic whereas the categories for classification are domain specific, there is a gap between the generation of object proposals and the classification of object proposals. In this paper, by taking advantages of the intrinsic structure and the complexity of each category of objects, we propose a novel tree-based hierarchical model to learn object proposals, from top proposals produced by the existing object proposals generation methods. First, we develop a tree-structured representation for each object to capture its hierarchical structure feature. Second, we propose a 23D compact feature vector to represent objects’ visual features. Third, we formulate a learning schema which evaluates the objectness of each proposal. Experiments demonstrate the significant improvement of the proposed approach over the state-of-the-art method in terms of object detection rate. An application is proposed based on this approach to help children learn and recognize objects by their visual appearances and their sub-parts structures.
Keywords: Hierarchical tree model, object proposals, object detection, learning
