Introduction
3D Gaussian Splatting (3DGS) has emerged as a transformative technique for real-time scene reconstruction. However, its performance is heavily dependent on the quality of its initial state. Typically, 3DGS initializes its Gaussians using a point cloud generated via Structure-from-Motion (SfM). In "few-shot" scenarios—where only a limited number of images are available—traditional SfM methods often fail to produce a sufficiently dense point cloud, leading to a "cold start" problem that degrades final reconstruction quality.
The Cold Start Problem in Few-Shot 3DGS
In few-shot settings, standard SfM tools like COLMAP are highly restrictive, often discarding points not visible in at least three views to ensure geometric accuracy. This results in a sparse, "holey" initial point cloud. During the 3DGS optimization phase, the system attempts to compensate by over-expanding the few available Gaussians to fill these gaps. This leads to the creation of "floaters" and significant blur, particularly in regions with weak textures or peripheral areas.
Methodology: Attentional Pointcloud Augmentation (APA)
EAP-GS addresses this bottleneck by focusing on initialization rather than complex optimization regularizations. The core of the method is Attentional Pointcloud Augmentation (APA), which proceeds in three stages:
- Detector-free SfM: The pipeline begins with a modern SfM approach that handles low-texture areas more robustly than traditional SIFT-based methods.
- Density Analysis: Using DBSCAN clustering, the system identifies "attention regions" where the current point cloud is sparse despite the presence of complex scene structures.
- Relaxed Triangulation: In these identified regions, the algorithm relaxes standard constraints to allow the generation of 3D points from two-view tracks. While these points are slightly less accurate than three-view tracks, they serve as vital geometric anchors in previously empty spaces.
Results and Performance
The EAP-GS method was evaluated on standard benchmarks including LLFF (3-view setting) and Mip-NeRF360 (12-view setting).
- Reconstruction Quality: On the LLFF dataset, EAP-GS achieved a PSNR of 18.93, a substantial improvement over the 14.63 PSNR of original 3DGS. It matches or exceeds other state-of-the-art few-shot methods like FSGS and CoR-GS.
- Efficiency: A key advantage of EAP-GS is its speed. On the LLFF benchmark, it reached convergence in just 1.75 minutes, significantly faster than FSGS (25.49 minutes) and CoR-GS (14.96 minutes).
- Visual Fidelity: The method excels at reconstructing surfaces with weak textures, such as white walls and ceilings, where traditional SfM typically fails to find matches.
Discussion and Conclusion
EAP-GS demonstrates that for few-shot 3DGS, a slightly noisier but significantly denser initial point cloud is preferable to a sparse, "perfect" one. The optimization process is capable of refining Gaussian positions if they are distributed effectively at the start. By prioritizing intelligent initialization through APA, EAP-GS provides a computationally efficient path to high-quality scene reconstruction from limited viewpoints. Future work may explore incorporating generative priors to handle even more extreme cases of 1 or 2 views.
Summary updated