Human-Directed Search of Three-Dimensional Mesh Models Based on Shape Similarity

As the popularity of three-dimensional (3D) geometric models has increased, so too has interest in methods to search 3D models based on their shape similarity. Two of the most difficult problems in finding shape similarity between 3D geometric models are a feature vector that succinctly describes the shape of the model and a method to compute the distance between a pair of feature vectors that reflects user preferences. We describe a human-directed 3D shape similarity search method that reflects, to some extent, the user preferences by using a learning classifier. Given an example 3D mesh model, the system first retrieves, as an initial guess, a set of models that are similar to the query based only on an unbiased mechanical measure of the distance (i.e., the Manhattan distance) between a set of feature vectors. The user then iteratively refines the query by tagging a subset of the retrieved models as being either similar or dissimilar. The system learns the user's preference using a learning classifier support vector machine (SVM), so that the distance values between the set of feature vectors are altered to reflect these preferences. Experimental results show that our method is capable of retrieving 3D models that better reflect the preferences of the user than the simple method using only the Manhattan distance.