Skeleton Embedded Views for Non-Rigid 3D Shape Classification and Retrieval

With the application of deep learning in 3D recognition, view-based methods show excellent performance in non-rigid 3D model classification and retrieval. However, view-based methods often have problems such as occlusion and lack of spatial pose information which are caused by different viewpoints. Aiming at this problem, we propose a classification and retrieval method for non-rigid 3D model based on skeleton feature embedded view. Firstly, this paper uses virtual cameras to obtain multi-views of non-rigid 3D models from different viewpoints, and then extracts the skeleton features of the views by using the method of topology thinning, and embeds the skeleton features into the original views. Secondly, the network structure of ResNet50 is improved and used as the backbone network to build a multi-branch convolutional neural network (MBCNN); Finally, the classification and retrieval of non-rigid 3D models are realized by MBCNN. Experiments on the non-rigid 3D model datasets SHREC10, SHREC11 and SHREC15 demonstrate the effectiveness of the proposed algorithm. The classification accuracy rates for the above three datasets are 100%, 100%, and 99.6% respectively. The results show that our method is much better than the existing non-rigid 3D shape retrieval methods.