DiGS: Depth-Initialized Gaussian Splatting for Single-Object Reconstruction

Area of Science:

• Computer Vision
• 3D Graphics
• Geometric Modeling

Background:

• Gaussian Splatting is a leading method for high-fidelity 3D scene reconstruction.
• Initialization strategies critically impact the convergence and quality of 3D reconstruction.
• Existing methods may struggle with fine details and initial optimization phases.

Purpose of the Study:

• To investigate the impact of initialization strategies on Gaussian Splatting.
• To introduce and evaluate a novel depth-initialized approach for improved 3D reconstruction.
• To enhance the recovery of fine-grained details and accelerate early optimization.

Main Methods:

• Developed Depth-initialized Gaussian Splatting (DiGS) pipeline.
• Utilized depth data from calibrated stereo cameras for initialization.
• Performed evaluations using synthetic and real-world datasets.
• Conducted objective metric analysis and a 20-participant user study.

Main Results:

• Depth-based initialization significantly enhances early optimization stages.
• DiGS demonstrates superior performance in reconstructing isolated single objects.
• Improved recovery of fine-grained details compared to standard methods.
• User study confirmed enhanced perceptual quality.

Conclusions:

• Initialization strategy is crucial for effective Gaussian Splatting.
• DiGS offers a robust and efficient method for high-quality 3D reconstruction.
• The approach shows promise for applications requiring detailed object modeling.