Using Multi-scale Gaussian Derivatives for Appearance-Based Recognition

This paper addresses a novel global appearance-based approach to recognize objects in images by using multi-scale Gaussian derivatives (GDs). Because the GDs' distributions of filtered images are almost picked, this obstacles obtaining discriminative binned distributions for each image. For this reason, we execute k-means clustering on each scale of pooled Gaussian derivative set of the instances come from all classes to yield k-cluster centroids for partitioning feature space, thus generating normalized binned marginal distributions for all training and testing samples, which are holistically adaptive to underlying distributions. On similarity matching, we identify each image with a point of product multinomial manifold with boundary, and use the direct sum of geodesic distance metric for sets of binned marginal densities. The promising experimental results on Zurich buildings database (ZuBuD) validate the feasibility and effectiveness of our approach.