Efficiency Investigation of BoF, SVT and Pyramid Match Algorithms in Practical Recognition Applications

The choice of local image features is crucial for many computer vision applications. Scale Invariant Feature Transform (SIFT) features [1] and their upgraded version - Speeded-Up Robust Features (SURF) [2], are the most successful and popular ones for different object and scene recognition tasks, in terms of non-real and real time requirements, respectively. However, local features are not the only means building up the potential for fast and user-friendly solutions. Methods applied to process extracted features and their descriptors at the next steps are also critical. Three selected approaches of these type, based on to Bag of Features [3], Scalable Vocabulary Tree [4] and Pyramid Match [5] methods, respectively, are examined in the paper. Their effectiveness with regard to real-time make and model recognition of cars as well as visual building and places identification is reported and discussed as a final result of performed examination.