Yao, GrekouMavromatis, SebastienMari, Jean-LucLiu, LingjieAverkiou, Melinos2024-04-302024-04-302024978-3-03868-239-41017-4656https://doi.org/10.2312/egp.20241045https://diglib.eg.org/handle/10.2312/egp20241045Recent progress in 3D reconstruction has been driven by generative models, moving from traditional multi-view dependence to single-image diffusion model based techniques. However, these innovative approaches often face challenges with sparse view scenarios, requiring known poses or template shapes, often failing in high-resolution reconstructions. Addressing these issues, we introduce the ''F2F'' (Few to Full) framework, designed for crafting high-resolution 3D models from few views and unknown camera poses, creating fully realistic 3D objects without external constraints. F2F employs a hybrid approach, optimizing both implicit and explicit representations through a unique pipeline involving a pretrained diffusion model for pose estimation, a deformable tetrahedra grid for feature volume construction, and an MLP (neural network) for surface optimization. Our method sets a new standard by ensuring surface geometry, topology, and semantic consistency through differentiable rendering, aiming for a comprehensive solution in 3D reconstruction from sparse views.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Sparse views; 3D reconstruction; Hybrid 3D representation; Differentiable renderingComputing methodologies → Sparse views3D reconstructionHybrid 3D representationDifferentiable rendering10.2312/egp.202410452 pages