A vca-based approach to enhance learning data sets for object classification

Abstract

This paper presents a novel approach to solving the problem of poor learning data in complex object classification task. It efficiently combines the Visual Content Analysis technique known as the Scalable Vocabulary Tree (SVT) and contour-based descriptors to recommend new training samples. The SVT technique uses the SIFT features to identify and accurately localize objects of interest within the visual content of the processed query images. Despite the small learning data set its classification accuracy is pretty good and matches the accuracy of a dedicated CNN network trained under the same conditions. However, due to the ability of fast and effective incremental learning, it overcomes the convnet type networks. Contour-based classification based on Point Distance Histogram (PDH) is utilized then to increase the classification certainty. During this stage, the PDH descriptors representing a given object of interest are matched against descriptors stored in the pattern database, where each object is represented by a collection of 360 pattern outlines extracted from its 3D model. As finally reported, such an exact pattern representation allows for achieving a high classification accuracy of the entire approach.