Clustering based model datasets visualization and retrieval

We present in this paper a clustering-based visualization method for 3D model datasets. Firstly, Isometric feature mapping (Isomap) algorithm is used to reduce high-dimensional data of 3D model to three dimensional data. The reduced data is then used to learn the cluster representative models. Then Particle Swarm Optimization (PSO) is introduced to calculate the geometric median of a model cluster, and the data point which is closest to the geometric median of the cluster is selected as the representative of this cluster. Finally, combining with model alignment approaches, the orientation of the representative model is determined. Furthermore, according to the similarity between a query model and cluster representatives, a process of model retrieval is proposed. The first step of this process is to find the representative models which are most similar to the query model. The search is then restricted within the corresponding clusters which decreases quantity of candidate models. Our experimental results demonstrate that this process can achieve a substantial increase in retrieval efficiency without any loss in retrieval accuracy if an appropriate parameter combination is used.