Niloufar Salehi Dastjerdi and M. Omair Ahmad, Concordia University, Canada

Image descriptors play an important role in any computer vision system e.g. object recognition and tracking. Effective representation of an image is challenging due to significant appearance changes, viewpoint shifts, lighting variations and varied object poses. These challenges have led to the development of several features and their representations. Spatiogram and region covariance are two excellent image descriptors which are widely used in the field of computer vision. Spatiogram is a generalization of the histogram and contains some moments upon the coordinates of the pixels corresponding to each bin. Spatiogram captures richer appearance information as it computes not only information about the range of the function like histograms, also information about the (spatial) domain.However, there is a drawback that multi modal spatial patterns cannot be well modelled.Region covariance descriptor provides a compact and natural way of fusing different visual features inside a region of interest. However, it is based on a global distribution of pixel features inside a region and loses the local structure.In this paper, we aim toovercome the existing drawbacks of these descriptors. To this, we propose rspatiogram and then a new hybrid descriptor is presented which is combination of rspatiogram and traditional region covariance descriptors. The results show that our descriptors have the discriminative capability improved in comparison with other descriptors.

Feature Descriptor, Spatiogram, Region Covariance