Non-Iterative Normalized Feature Extraction In Large Viewpoint Variances Based On Pca Of Gradient

Effective local feature extraction is one of the fundamental tools for retrieval applications in computer vision. However, it is difficult to achieve distinguishable local features in large viewpoint variances. In this paper, we propose a novel non-iterative approach of normalized feature extraction in large viewpoint variances, which adapts local regions to rotation, scale variance and rigid distortion from affine transformation. Our approach is based on two key ideas: 1) Localization and scale selection can be directly achieved with the centroid and covariance matrix of the spatial distribution of pixels in a local region. 2) Principal Component Analysis (PCA) on gradients of intensity gives information on texture, thus it can be used to get a resampled region which is isotropic in terms of variance of gradient. Experiments demonstrate that our normalized approach has significant improvement on matching score in large viewpoint variances.