Different from previous local smoothing filters based on local geometry signal, a novel denoising technique based on a non-local averaging of geometry signal of all sampled points on the point set surface is proposed. By using the bilateral filtering operator, the differential signal for each discrete point is obtained. The final geometry information of sample point can be reconstructed as the averaged geometry gray level computed by the NL-means. The mixture tree is applied to accelerate the similarity matching computation, which makes the NL-means more efficient for dealing with large point set surface. Experimental results illustrate that the approach is efficient and satisfied.