AttentionGS: Towards Initialization-Free 3D Gaussian Splatting via Structural Attention
This addresses a robustness issue for 3D reconstruction in real-world applications, though it is incremental as it builds on existing 3D Gaussian Splatting methods.
The paper tackles the problem of 3D Gaussian Splatting's reliance on high-quality initial point clouds, which limits its use in texture-deficient or constrained-view scenarios, by proposing AttentionGS, a framework that uses structural attention for direct reconstruction from random initialization, achieving significant outperformance over state-of-the-art methods in unreliable initialization scenarios.
3D Gaussian Splatting (3DGS) is a powerful alternative to Neural Radiance Fields (NeRF), excelling in complex scene reconstruction and efficient rendering. However, it relies on high-quality point clouds from Structure-from-Motion (SfM), limiting its applicability. SfM also fails in texture-deficient or constrained-view scenarios, causing severe degradation in 3DGS reconstruction. To address this limitation, we propose AttentionGS, a novel framework that eliminates the dependency on high-quality initial point clouds by leveraging structural attention for direct 3D reconstruction from randomly initialization. In the early training stage, we introduce geometric attention to rapidly recover the global scene structure. As training progresses, we incorporate texture attention to refine fine-grained details and enhance rendering quality. Furthermore, we employ opacity-weighted gradients to guide Gaussian densification, leading to improved surface reconstruction. Extensive experiments on multiple benchmark datasets demonstrate that AttentionGS significantly outperforms state-of-the-art methods, particularly in scenarios where point cloud initialization is unreliable. Our approach paves the way for more robust and flexible 3D Gaussian Splatting in real-world applications.