Efficient modeling by selecting learning samples in human pose estimation

While the more learning data the better the recognition, increase in the data causes an expensive computational cost in learning. This paper proposes how to decrease the computational cost by appropriately selecting the learning data. In particular, we put our focus on learning for human pose estimation in still images. Three kinds of methods are proposed for learning data selection in this paper. The first one divides all data into several clusters in a feature space for avoiding duplication of similar data. The second one selects the data based on their distance from a discriminant plane for efficiently updating it. Third one merges those two methods as well as pruning in optimized pose search. Experimental results show that the proposed method can decrease the learning time by 79 % with less decrease in pose estimation accuracy.