3D Super-Resolution Reconstruction Based on Multi-view Representation

Despite the progressed achievement on three-dimensional generation, the results of the research on high-resolution models are not satisfactory. In order to generate 3D objects in high resolution, we implement a 3D model super-resolution reconstruction system. First, a set of orthogonal depth maps are acquired based on each input 3D model, which has a low resolution. Then we upscale these images by training a super-resolution network. The high-resolution orthogonal depth maps we obtained include rich information and are used to carve models in high resolution. Absolutely, our system is a reconstruction process from images to high quality 3D objects in voxel space. Considering the effect of the resolution of the image on the reconstruction results, we design a suitable super-resolution network to obtain a high-resolution depth map. We have conducted experiment on a subset of ShapeNet dataset and the experimental results show the effectiveness of our method.