Abstract: By automatically selecting a good viewpoint in the 3D visualization systems,users can acquire information of the 3D models more effectively.Based on the theory of information entropy and human perception,this paper presents an algorithm for automatic viewpoint selection.We design a perceptual feature function of the object with respect to all possible viewpoints.This function is then optimized by using a stochastic gradient descent like method.Experimental results show that our selected viewpoint provides a best view of the object which conforms very well with human perception.By using GPU,its time cost is also greatly reduced.