This paper Describes a method of selecting proper view direction for recognition of polyhedral objects. Our method is based on range maps and have two simple criteria: (a) the number of visible planes, (b) the uniformity of areas of image planes, to evaluate images taken from some view directions. Gradient distribution which represents the histogram of gradients have some information about feature of planes, so that we found that evaluation based the criteria correspond to entropy of gradient distribution. Several experiments, using real range maps, are presented and the efficiency of this entropy measure is discussed.