This paper describes a new 3D model retrieval approach based on shape distributions, which represents a 3D shape as three D2 shape histograms. First, generate a sufficiently large number of random sample points on surface of 3D model and note the normal of each point. Second, we adopt the D2 shape function to measures the distance between two random points on the surface of the model, at the same times, compute the angles between the normals of the points and the line segment which is formed by the two random points. Last, according to the angles, we sort the line segments into three sets. For each set, we calculate shape distribution histograms associated with the D2 shape distribution function. Compare each pair of shape distribution histograms using well-known curve matching techniques and add the three weighted results. The sum is regarded as the similar coefficient of the two models. Experimentation shows this algorithm can efficiently give the similar degree of the two models and as well reflecting human perceptual similarity.