Objects classification in still images using the region covariance descriptor

Abstract

The goal of the Object Classification is to classify the objects in images. Classification aims for the recognition of generic classes, which is also known as Generic Object Recognition. This is quite different from Specific Object Recognition, such as recognizing specific person, own car, and etc. Human beings are generally better in recognizing generic classes than specific objects. Classification is a much harder problem to solve by artificial systems. Classification algorithm must be robust to changes in illumination, object scale, view point, and etc. The algorithm also has to manage large intra class variations and small inter class variations. In recent literature, some of the classification methods use Bag of Visual Words model. In this work the main emphasis is on region descriptor and representation of training images. Given a set of training images, interest points are detected through interest point detectors. Region around an interest point is described by a descriptor. Region covariance descriptor is adopted from porikli et al. [21], where they used this descriptor for object detection and classification. This region covariance descriptor is combined with Bag of Visual words model. We have used a different set of features for Classification task. Covariance of d-features, e.g. spatial location, Gaussian kernel with three different s values, first order Gaussian derivatives with two different s values, and second order Gaussian derivatives with four different s values, characterizes a region of interest. An image is also represented by Bag of Visual words obtained with both SIFT and Covariance descriptors. We worked on five datasets; Caltech-4, Caltech-3, Animal, Caltech-10, and Flower (17 classes), with first four taken from Caltech-256 and Caltech-101 datasets. Many researchers used Caltech-4 dataset for object classification task. The region covariance descriptor is outperforming SIFT descriptor on both Caltech-4 and Caltech-3 datasets, whereas Combined representation (SIFT + Covariance) is outperforming both SIFT and Covariance