A center surround contrast based feature for object pose estimation

We propose a primary visual cortex inspired oriented edge feature for object pose estimation. The neural feedback like feature is based on a Center-Surround Contrast excitation and a k-Winner-Take-AU inhibition, to extract different orientations of edge response from an image patch. To compute local descriptor, we model each oriented edge response with a PDF distribution, before concatenating their attributes from all orientations. To choose a suitable PDF candidate during training, we ran a similarity test fit between empirical and parametric statistics. We train a bank of binary view pose classifiers using SVM on dense features with Spatial Pyramid Representation [15]. We evaluate and compare the Mean Average Precision of our proposed descriptor with HOG [6] for pose estimation evaluation. Lastly, we showed that using our proposed feature over baseline resulted in a gain of nearly 15% on the EPFL Multi-View Car Dataset [2].