Research On Panoramic-view-based 3D Model Classification And Retrieval

With the development of computer vision,the application of 3D models in this direction has become more and more extensive.It is a problem how to classify and find the best matched model efficiently.There are many methods about 3D models classification and retrieval,and the view-based image representation method has good performance in this field,and has achieved good results,not only on databases but also in competitions.In this paper,we introduce the representation of 3D model based on panorama,with the method of multi-views representation of 3D model.The main introductions are as follows:(1)Extract the panorama of 3D models.Place the gridded 3D model in a fixed cylinder and map the distance from the main axis of the 3D model to the surface and the orientation of the 3D patch to the side of the cylinder,and then expand the side of the cylinder to form a rectangular panorama.In this way,the distance from the surface of the 3D model to the main axis and the orientation of the surface patches are more fully considered with the structural features of the 3D model.(2)we proposed a new network,in order to extract the multi-scale global and local features of all panoramas,to generate a powerful feature descriptor.With the different scale panoramas,our network gets the depth features of different scales,and combine the depth features of different scales through multiple channels to generate the final feature descriptor of the 3D models.The local and global features of the panorama can be obtained by different scaled panorama.By combining the multi-channel features,the connections between different panoramas can be obtained.(3)The attention mechanism is added to the proposed network to quantize the contribution of the features outputted by different channels to the final descriptor.In the end,we experiment our methods in the 3D model databases Model Net10,Model Net40 and Shape Net Core55.The results show that our methods have good performance.