GaussianObject: High-Quality 3D Object Reconstruction from Four Views with Gaussian Splatting

TL;DR

GaussianObject reconstructs a high-quality 3D Gaussian representation from only four images by combining explicit geometric initialization, floater removal, and a diffusion-based repair model for unseen content.

Abstract

Reconstructing a 3D object from very sparse views is difficult because few images provide weak multi-view constraints and leave large areas unseen. GaussianObject reconstructs and renders an object from only four input images using 3D Gaussian splatting. Visual-hull initialization and floater elimination inject structural priors into early optimization, producing a coarse but consistent Gaussian representation. A diffusion-based Gaussian repair model then supplements missing object information and refines the representation; its training pairs are obtained through a self-generating strategy. A COLMAP-free variant removes the need for precomputed accurate camera poses while retaining competitive quality. Evaluations on MipNeRF360, OmniObject3D, OpenIllumination, and collected unposed images show strong rendering and reconstruction performance under extreme view sparsity.

Key contributions

• Reconstructs high-quality 3D Gaussian objects from only four views.
• Combines structure-aware initialization with diffusion-based repair of unseen content.
• Provides a COLMAP-free variant for reconstruction without accurate input poses.

Citation

@article{yang2024gaussianobject,
  title={GaussianObject: High-Quality 3D Object Reconstruction from Four Views with Gaussian Splatting},
  author={Yang, Chen and Li, Sikuang and Fang, Jiemin and Liang, Ruofan and Xie, Lingxi and Zhang, Xiaopeng and Shen, Wei and Tian, Qi},
  journal={ACM Transactions on Graphics},
  year={2024},
  doi={10.1145/3687759}
}