Joint local experts for measuring objectness of image proposal windows

In order to improve the efficiency and accuracy of object detection, the joint local experts method is proposed to estimate the objective windows by measuring how likely it is for an image proposal window to contain an object. Firstly, the proposal windows that do not contain any object obviously are filtered out by the local expert inter-union set. Then, the rest proposal windows that contain the object are measured by local expert cosine similarity. Finally, the objective windows are estimated by local expert non-maximum suppression from a large number of proposal windows that repeatedly contain the same object. Experiment results show that the proposed method is able to efficiently estimate the objective windows which accurately contain the object.