View-independent object detection using shared local features

In this study, we developed a novel method for detecting view-independent objects in a cluttered background with partial occlusion using shared features. These shared features are selected as common features among classes while the detectors used for each class are trained jointly rather than independently using shared features, which reduces the number of classifiers. We developed an exhaustive greedy selection method for selecting shared features and training their classifiers using only the shared features. The exhaustive greedy selection method randomly selects an exhaustive set of rectangular local features in a normalized object window and selects n significant shared local features from 12 different viewpoints and their effective shared classifiers using random forests. An integral histogram based on oriented-center symmetric local binary pattern (OCS-LBP) descriptor is used to represent a shared feature and to reduce the feature dimensions effectively. The final score is summed bilinearly using the probabilities of neighboring views to determine the location and viewpoint of the object because each view overlaps with neighboring views. Our proposed algorithm was successfully applied to the PASCAL VOC 2012 dataset and its detection performance was better than other methods.