Object Recognition Using Words Model of Optimal Size in Histograms of Oriented Gradients

This paper presents a new bag-of-words based algorithm for object recognition. Our algorithm also includes five steps: feature detection and representation, codebook generation, learning and recognition. All features are extracted as dense grids of images instead of interest point for computationally efficiency and effectiveness. While features are described by histograms of oriented gradients (HOG) rather than widely used scale invariant feature transform (SIFT). Different from previous work that randomly determine codebook size, the influence of the codebook size on the performance is also explored as another key problem. Then optimum sizes are chosen for further operation. In the training part, linear support vector machine (SVM) as well as support vector regression (SVR) is adopted due to their sati. Experiment are done on various cars from one of the most challenging databases Gaze; and the results show that our algorithm achieves desirable effect while the computational cost hardly increases.