Point cloud models (PCMs) are 3D point datasets from 3D acquisition device. They are densely sampled from the surfaces of objects. Each point in the PCM consists only of 3D coordinates, sometimes also of normal vector without information of structures, or connectivity. The size of a PCM may be up to several millions. Research reported in this thesis focuses on developing a two-stage technique: to simplify a very dense PCM into compact PCMs according to required compact rate, and on reverse, to refine the compact PCMs at least as dense as or denser than the original.  In stage one, a PCM (only including 3D coordinates), will be simplified into a compact one with an octree and principle component analysis (PCA). From the result of PCA, features of the local surface defined by points in a leaf node can be detected. For a specific feature in a leaf node of the octree, corresponding simplification algorithm is applied to resample points from original one. The points in compact PCM also have information of normal vector and feature . On stage two, a dense PCM is obtained by refining the compact PCM from stage one with the refinement schemes. Finally, in order to check the results of simplification and refinement, we make two comparisons. The first is between compact PCM and the original PCM in order to determine if it is good or bad for a compact PCM to represent the original PCM . The second comparison happens between refined and original PCMs in order to verify how many features are lost during the two stages, and for survived features, how different they are from the ones in the original PCM. For both comparisons, a normalized result is reported.