| Object detection in 2D images remains a key challenge in computer vision. There are many different machine learning and feature extraction algorithms available. As there is no free lunch, we have no complete understanding of what method would work in what context. To that end, we perform a comparative analysis of different detection methods and different feature vectors, in this study. In this work we are concentrating on the state of the art machine learning algorithms Adaboost, Random Forest, Random Ferns and Support Vector Machine. The features we evaluate are two variants of histogram of oriented gradients, Haar-like features and raw pixel values. For experiments we have used the PASCAL VOC 2007 benchmark dataset, which has 23 classes. We have performed detailed analysis of performance of each feature and machine learning algorithm, as well as combinations of them over all 23 classes of the benchmark dataset. We have also compared the performance of each feature-learning algorithm combination on each object class. Our results indicate that the highest performing combination over all object classes is Adaboost with Haar-like features. Random Forest with Haar-like features performed almost as well. We also determined that the best features among the provided set of features to use with linear SVM are histogram of oriented gradients. |
