Improved HOG Descriptors

We study the feature set for object recognition problem, and use human detection as a test case. We propose two improvements based on HOG model which are Spatial Selective Method and Multi-level Method. In Spatial Selective One, we use HOG descriptor to extract feature vector from image window, but we shorten the feature vector size by eliminating less informative region. We get the same performance as Dalal´s method, while reducing the extraction running time by 40%. In the Multi-level Method, we enhance the performance of HOG descriptor by 3% by adding more information to feature vector set through using concatenating multi-level on extraction process. All the experiments of this work are evaluated on INRIA pedestrian dataset 2009.