FreeSplat++: Generalizable 3D Gaussian Splatting for Efficient Indoor Scene Reconstruction

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Existing 3D Gaussian Splatting (3DGS) methods struggle with large-scale whole-scene reconstruction quality and efficiency.
• Current approaches often focus on limited regions, failing to generalize to extensive indoor environments.

Purpose of the Study:

• To develop an advanced generalizable 3DGS method, FreeSplat++, for efficient and accurate large-scale indoor whole-scene reconstruction.
• To significantly accelerate reconstruction speed and enhance geometric accuracy in complex scenes.

Main Methods:

• Proposed a Low-cost Cross-View Aggregation framework for processing long input sequences in whole-scene reconstruction.
• Introduced a pixel-wise triplet fusion method to incrementally aggregate and reduce redundancy of 3D Gaussian primitives.
• Implemented a weighted floater removal strategy and a depth-regularized per-scene fine-tuning process for enhanced accuracy and quality.

Main Results:

• FreeSplat++ demonstrates superior performance in whole-scene reconstruction compared to existing generalizable 3DGS methods.
• Achieved substantial improvements in reconstruction accuracy and a notable reduction in training time compared to conventional per-scene optimized 3DGS.

Conclusions:

• FreeSplat++ offers a viable alternative for large-scale indoor whole-scene reconstruction, balancing quality, efficiency, and accuracy.
• The proposed methods effectively address limitations of previous generalizable 3DGS techniques in complex environments.