An Improved 3D Shape Recognition Method Based on Panoramic View

Recognition of three-dimensional (3D) shape is a remarkable subject in computer vision systems, because of the lack of excellent shape representations. With the development of 2.5D depth sensors, shape recognition is becoming more important in practical applications. Many methods have been proposed to preprocess 3D shapes, in order to get available input data. A common approach employs convolutional neural networks (CNNs), which have become a powerful tool to solve many problems in the field of computer vision. DeepPano, a variant of CNN, converts each 3D shape into a panoramic view and shows excellent performance. It is worth paying attention to the fact that both serious information loss and redundancy exist in the processing of DeepPano, which limits further improvement of its performance. In this work, we propose a more effective method to preprocess 3D shapes also based on a panoramic view, similar to DeepPano. We introduce a novel method to expand the training set and optimize the architecture of the network. The experimental results show that our approach outperforms DeepPano and can deal with more complex 3D shape recognition problems with a higher diversity of target orientation.