RealImage2Mesh: Reconstructing symmetrical mesh models from single real-world images

We propose a method to reconstruct mesh models from single real-world images. Our method takes a real-world image of an object as input, uses an encoder-decoder network to extract 2. 5D information such as the silhouette, depth, and normals of the object, and then inputs this information into a multi-level 3D mesh generation network to generate a high-quality mesh model using mesh deformation and face pruning. In the process of generating the mesh model, we propose a symmetry loss to keep the shape of the mesh model reasonable. According to our experiments, compared with other single-view object reconstruction methods, our method can reconstruct 3D models with higher quality from real-world images.