摘要
Reconstructing high-fidelity 3D geometry and appearance from sparse-view images is challenging when reliable camera poses are unavailable. Existing sparse-view methods typically assume known intrinsics/extrinsics or rely on explicit pose estimation, which reduces practical applicability and increases pipeline complexity. Moreover, mesh colorization is often handled as a separate task, despite per-vertex color being a standard representation of surface-aligned attributes in scientific and engineering workflows. We propose TNF-US, a unified pose-free framework that predicts a triplane neural field directly from a sparse set of RGB images, enabling both NeRF-style novel-view synthesis and mesh colorization in a single feed-forward model. Cross-view dense correspondences are distilled into per-view geometric descriptors that modulate ViT image tokens. The conditioned multi-view tokens are fused and decoded into triplanes that act as a shared latent for volumetric rendering and direct vertex queries. Training combines photometric supervision with vertex-color regression and a point-sampled density regularizer to stabilize geometry and encourage surface-consistent appearance. Experiments on Google Scanned Objects and OmniObject3D demonstrate consistent improvements over strong baselines, achieving up to 6.7% PSNR gain and 47.8% LPIPS reduction on novel-view evaluation, and up to 72% lower vertex-color error for mesh colorization.
| 源语言 | 英语 |
|---|---|
| 期刊 | IEEE Transactions on Visualization and Computer Graphics |
| DOI | |
| 出版状态 | 已接受/待刊 - 2026 |
| 已对外发布 | 是 |
学术指纹
探究 'Triplane Neural Fields from Unposed Sparse-Views for 3D Reconstruction and Mesh Colorization' 的科研主题。它们共同构成独一无二的学术指纹。引用此
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver