Learning Optimal Semantically Relevant Visual Dictionary using Evolutionary Algorithms
We propose a novel maximally discriminative visual dictionary learning method using evolutionary algorithms. Due to large intra-category appearance variation, semantically equivalent descriptors are scattered over feature space . Current methods, using hard or soft assignment, are unable to cluster these equivalent descriptors, inter-leaved with other categories, together. We over partition feature space, generating a huge population of tiny syntactically cohesive partitions, which are optimally assigned, with locality constraints, so as to obtain a maximally discriminative visual dictionary.