Gamba: Marry Gaussian Splatting With Mamba for Single-View 3D Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Single-view 3D reconstruction is computationally intensive.
• Existing methods often require extensive training or optimization.
• Efficient and accurate 3D asset generation from limited input is a key challenge.

Purpose of the Study:

• To develop an end-to-end 3D reconstruction model capable of millisecond-speed generation.
• To introduce a novel architecture for efficient 3D Gaussian Splatting (3DGS) reconstruction.
• To enhance robustness by eliminating the need for 3D point cloud warmup supervision.

Main Methods:

• Introduced GambaFormer, a Mamba-based network for sequential 3DGS prediction with linear scalability.
• Developed radial mask constraints derived from multi-view masks for robust training.
• Trained the model on the Objaverse dataset and evaluated on the GSO Dataset.

Main Results:

• Gamba achieves end-to-end single-view 3D reconstruction using 3DGS.
• Reconstruction is completed in 0.05 seconds on a single NVIDIA A100 GPU, approximately 1,000x faster than optimization-based methods.
• Demonstrated competitive qualitative and quantitative generation capabilities compared to existing approaches.

Conclusions:

• Gamba presents a significant advancement in efficient and fast single-view 3D reconstruction.
• The Mamba-based architecture and novel constraints enable unprecedented speed and accuracy.
• This work paves the way for real-time 3D asset generation applications.