A geometric compression algorithm is presented in this paper to save the geometry model storage and transmission time. This method decomposes the model into a series of layers in a way of region growing. These layers are then encoded effectively by using inter-layers coherence, so that the topology information of the model is compressed dramatically. Experimental results show that it takes only an average of 1.42 bits per triangle. Additionally, a non-linear geometry predictor is designed to compress the geometric information. Compared with the previous work, this algorithm is of linear complexity, and it can be implemented effectively.