GaussianObject logo

GaussianObject

Paid

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

5.0
Inputs: image
Type
Saas

About GaussianObject

GaussianObject is a research framework for high-quality 3D object reconstruction from only four input images using Gaussian splatting. It addresses challenges of multi-view consistency and incomplete object information from sparse views by introducing visual hull initialization and floater elimination as structure priors, followed by a Gaussian repair model based on diffusion models that refines the representation. The framework supports both COLMAP-based and COLMAP-free (no pre-given camera poses) settings. Evaluated on MipNeRF360, OmniObject3D, OpenIllumination, and unposed image datasets, GaussianObject outperforms previous state-of-the-art methods. Published at SIGGRAPH Asia 2024 in ACM Transactions on Graphics.

Key Features

Visual hull initialization for multi-view consistency
Floater elimination to refine 3D Gaussians
Gaussian repair model based on diffusion models
Self-generating leave-one-out strategy for training repair model
COLMAP-free variant without pre-given accurate camera poses
Achieves high rendering quality from only 4 input images

Pros & Cons

Pros
  • High rendering quality from only four input images
  • Works without accurate camera poses (COLMAP-free)
  • Outperforms previous state-of-the-art methods on multiple challenging datasets
Cons
  • Requires at least four input images (not fewer)
  • Repair model training involves a self-generating strategy that adds complexity
  • Currently a research framework, not a ready-to-use commercial product

Best For

3D object reconstruction from highly sparse viewsApplications in 3D vision where only a few images are availableReconstruction from unposed images (COLMAP-free setting)

Alternatives to GaussianObject

FAQ

How many input images does GaussianObject need?
GaussianObject achieves high-quality 3D reconstruction using only four input images.
Does GaussianObject require camera poses?
It offers a COLMAP-free variant that does not require pre-given accurate camera poses, while also supporting traditional COLMAP-based settings.
What datasets were used for evaluation?
The method was evaluated on MipNeRF360, OmniObject3D, OpenIllumination, and unposed image datasets.
Is GaussianObject published?
Yes, it was presented at SIGGRAPH Asia 2024 and published in ACM Transactions on Graphics.