GaussianObject achieves high rendering quality, works with sparse views, and does not require pre-given accurate camera poses.
A technique for representing and rendering 3D objects with Gaussian splatting.
Techniques for building multi-view consistency and refining the initial optimization process.
A model for supplementing omitted object information and refining Gaussians.
A variant that does not require pre-given accurate camera poses, facilitating wider applications.
A strategy for obtaining image pairs for training the repair model.
Reconstructing and rendering 3D objects from highly sparse views.
Promoting applications of 3D vision techniques and improving user experience.
Achieving high rendering quality with only 4 input images.
Initialize 3D Gaussians using visual hull and floater elimination.
Train the Gaussian repair model using the self-generating strategy.
Use the trained model to correct views that need to be rectified.
Optimize 3D Gaussians using the corrected views and reference images.