Novel Gabor filter-based patch descriptor

One of the most important tasks of component-based object detection is locating the characteristic object parts and describing them in order to be easily localizable. This paper presents a local image descriptor based on Gabor wavelets. Due to the high dimensionality of the parameter space, these functions can be defined only laboriously, and it is very difficult to integrate them into the desirable application. At first, in order to reduce the number of parameters, some theoretical relations are taken into account. Secondly, in the experiment phase, the most adequate filters for a given image patch are determined with the GentleBoost learning algorithm. Using this descriptor, a responses-map is created, which can be easily combined with the well-known deformable object model in order to detect the desired parts of the target object. The advantage of our descriptor compared to those found in the state of the art is that the selected filter set is not always the same, but after fine-tuning parameters, it selects the most appropriate filters for the target patch so that the descriptor becomes general, and at the same time, discriminative as well.