Effective object proposals: size prediction for pedestrian detection in surveillance videos

Though detectors based on deep networks have witnessed a significant boost in pedestrian detection performance, small-scale pedestrian detection remains to be a challenging task. To this end, the authors propose a scale prediction method to eliminate the dependence of most existing detectors on pre-defined anchor boxes. Due to projective transformation, a pedestrian standing afar will appear smaller than one standing closeby. For a stationary surveillance camera, different blocks on the image correspond to the views of various depths leadings to diverse sizes of object proposals. A neural network structure is developed to empirically estimate the size of a pedestrian candidate in the pixel coordinates given its central location. Comprehensive sets of experiments on two real-world datasets demonstrate that the proposed method achieves superior performance with state-of-the-art methods. Besides, significant performance advantages on small pedestrian detection are observed.