Weakly supervised learning of component-based hierarchical model for object detection

In this paper, we present a hierarchical framework for detecting and localizing object by components. The system is structured with a root detector and several component detectors that are trained to separately find the object and different parts of the object on the first level. On the second level the spatial relations model performs detection by combining the root detector and the component detectors. We learn the component models in a weakly supervised manner, where object labels are provided but component labels are not. The root model and each component model are learned by using boosting. The weak classifiers are vector-valued HOG features which are projected from d-dimensional to 1-dimensional subspace by Fischer Linear Discriminant (FLD). The experimental results demonstrate that our method is comparable with the previous ones.