Monocular 3D Object Reconstruction with GAN Inversion

Recovering a textured 3D mesh from monocular image is highly challenging, particularly for in-the-wild objects that lack ground truths. In this work, we present MeshInversion, novel framework to improve the reconstruction by exploiting generative prior of GAN pre-trained synthesis. Reconstruction achieved searching latent space in best resembles target accordance with single view observation. Since encapsulates rich semantics terms geometry and texture, within manifold thus naturally regularizes realness fidelity reconstruction. Importantly, such regularization directly applied space, providing crucial guidance parts are unobserved 2D space. Experiments on standard benchmarks show our obtains faithful reconstructions consistent texture across both observed parts. Moreover, it generalizes well meshes less commonly seen, as extended articulation deformable objects. Code released at https://github.com/junzhezhang/mesh-inversion .