Efficient Object Detection Based on Several Samples

Object detection by classifier is unavailable when there exist only several samples. In this case, the traditional solution is running the template matching method several times. The efficiency is decreased and the additional object information from several samples is not used. In this paper, a novel detection framework based on several samples is proposed to handle this case. Firstly, the cell sizes of voting space are trained based on several samples, which represent the tolerance of appearance at each pixel location, where the voting space is constructed by the image coordinates and the ranges of the feature value at the corresponding position. Next, in the detection part we propose a much more efficient strategy than that of overlapping in [12], which contains twice localizations. In the first localization, the target image divides into fixed-size blocks that are used to decide whether they are the components of the object. Then we use the sample size patches of target image, which are localized by the blocks, to execute the second localization of object. Compared with the previous template matching method (as LARK), our method needs smaller memory usage and has a better performance both in accuracy and efficiency.