3D Voxel Model Retrieval Based on Octree Structure

With the development of VR/AR technology and the wider 3D applications, it is realized that 3D model retrieval is becoming more and more important. Model-based retrieval preserves the spatial and geometric features, which includes not only the surface information but also the internal properties of the model. However, there are concerns in relate to its high storage and high computation. Deep learning has demonstrated successful breakthroughs in the fields of speech recognition, graphic image classification and natural language processing etc. In this paper, after studying the 3D model pre-processing and 3D model representation, a method combined with 3D voxel model and octree structure is proposed for 3D model retrieval. First of all, the coarse-grained features and fine-grained features of the voxelization model are extracted. After the fusion, the features which expressed in the form of octree are input into the convolutional neural network for training, and the Euclidean distance is the metric for evaluating and retrieving the model in the end. In order to form an octree for storing the 3D model, eight equal cubic meshes are able to be divided after the 3D model is scaled and aligned with a standard unit 3D boundary cubic volume. Such a mesh process will continue for each cubic volume, which include the 3D model, until the mesh quality reaches the requirement. By using the octree feature representation, not only the storage consumption is effective reduced due to the process of voxelization, the details of the original 3D mesh model are also preserved. The presented algorithm uses the improved Octree structure as the basic data structure of the model voxelization which is applied to the convolutional neural network for model classification. By designing a novel spatial octree, a 3D model is represented by which surface information was stored into the leaf nodes of the octree. The leaf nodes are able to be trained as initial data and evaluated through the improved octree neural network structure on GPU. IO-CNN is able to support various CNN structures with different 3D representations to extract and classify the 3D model for 3D model retrieval. With careful analyzing of 3D model, it is found that it is unnecessary to process interior part of the 3D model for voxelization if it is a closed 3D geometry. The voxelization of the interior part of the 3D model will never affect the representation of geometric features. In considering the computational performance, such modifications were made during the process of model voxelization. After voxelization, the normalized 3D model is represented by octree for obtaining the spatial information. An iterate process is carried out which 1 is set to the region which includes the 3D model and 0 is set to the region without the model respectively. By only voxelizing the outer surface of the model, the computational overhead is greatly reduced due to the optimized data storage and convolutional neural network training. The experiment has shown, with applying the SOFTMAX cost function, after a large amount of training data through convolutional neural network, the presented algorithm has more advance in 3D model retrieval than other similar algorithms. © 2021, Science Press. All right reserved.