ITERATIVE LOCALIZATION REFINEMENT IN CONVOLUTIONAL NEURAL NETWORKS FOR IMPROVED OBJECT DETECTION

Accurate region proposals are of importance to facilitate object localization in the existing convolutional neural network (CNN)-based object detection methods. This paper presents a novel iterative localization refinement (ILR) method which, undertaken at a mid-layer of a CNN architecture, iteratively refines region proposals in order to match as much ground-truth as possible. The search for the desired bounding box in each iteration is first formulated as a statistical hypothesis testing problem and then solved by a divide-and-conquer paradigm. The proposed ILR is not only data-driven, free of learning, but also compatible with a variety of CNNs. Furthermore, to reduce complexity, an approximate variant based on a refined sampling strategy using linear interpolation is addressed. Simulations show that the proposed method improves the main state-of-the-art works on the PASCAL VOC 2007 dataset.