EKF-GS: An Improved 3D Gaussian Splatting Using Extended Kalman Filter.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• 3D Gaussian Splatting (3DGS) is a rendering technique for real-time novel view synthesis.
• Existing 3DGS methods often require extensive training time and lack robust uncertainty estimation.
• Efficient optimization and uncertainty quantification are critical for advancing 3DGS.

Purpose of the Study:

• To develop a hybrid optimization framework for 3D Gaussian Splatting.
• To enhance convergence speed and reconstruction quality.
• To introduce uncertainty quantification and uncertainty-guided densification into the 3DGS process.

Main Methods:

• Integration of Extended Kalman Filter (EKF) with stochastic gradient descent (SGD) into a unified framework named EKF-GS.
• Development of an uncertainty-guided Gaussian densification strategy.
• Implementation of uncertainty quantification capabilities within the optimization pipeline.

Main Results:

• Achieved faster convergence compared to standard SGD-based 3DGS methods.
• Demonstrated improved 3D reconstruction quality with fewer training iterations.
• Showcased reduced overall training time on public benchmark datasets.
• Successfully implemented uncertainty quantification for Gaussian splats.

Conclusions:

• EKF-GS offers a significant advancement in 3D Gaussian Splatting optimization.
• The hybrid approach effectively balances convergence speed, reconstruction accuracy, and uncertainty estimation.
• This framework paves the way for more efficient and reliable 3D scene representation.