Sphere Face Model: A 3D morphable model with hypersphere manifold latent space using joint 2D/3D training

Abstract 3D morphable models (3DMMs) are generative for face shape and appearance. Recent works impose recognition constraints on 3DMM parameters so that the shapes of same person remain consistent. However, traditional 3DMMs satisfy multivariate Gaussian distribution. In contrast, identity embeddings meet hypersphere distribution, this conflict makes it challenging reconstruction to preserve faithfulness consistency simultaneously. other words, loss can not decrease jointly due their To address issue, we propose Sphere Face Model (SFM), a novel monocular reconstruction, preserving both fidelity consistency. The core our SFM is basis matrix which be used reconstruct shapes, basic learned by adopting two-stage training approach where 2D data in first second stages, respectively. We design resolve distribution mismatch, enforcing have hyperspherical Our model accepts constructing sphere models. Extensive experiments show has high representation ability clustering performance its parameter space. Moreover, produces high-fidelity consistently conditions reconstruction. code will released at https://github.com/a686432/SIR