An Effective Way of 3D Model Representation in Recognition System

Recognition of 3D objects is among the most popular topics in computer vision, and to find an effective representation of 3D models is a key issue. This paper proposes a novel way to describe 3D models in 3D object recognition system. We select three 2D shape features based on their complementarities, and implement feature fusion with coefficients obtained by self-learning method to form a concatenate feature with better robustness. Isomap manifold-learning-based clustering is introduced for more effective selection of representative views, because its non-linear property adapt to the 3D view sphere of objects very well, thus resulting in images with better representativeness. To test the effectiveness of this representation, a 3D object recognition system is established. Experiments on Princeton Shape Benchmark show the recognition rate of our method is comparative with state-of-the-art 3D model retrieval methods. The well-performed system can be a proof of the advancement of our method of 3D model representation.