The paper presents a comprehensive analysis of the flaw in the original adaptive density control strategy of 3D Gaussian Splatting (3D-GS), which leads to the issue of over-reconstruction and blurry rendering results. The authors identify the root cause as "gradient collision" - the pixel-wise sub-gradients of the view-space positional gradient may have different directions, causing them to cancel each other out during summation. This prevents the gradient magnitude from surpassing the densification threshold, hindering the split of large Gaussians in over-reconstructed regions.
To address this issue, the authors propose the novel "homodirectional view-space positional gradient" as the criterion for densification. By taking the absolute value of each gradient component before summation, the homodirectional gradient effectively eliminates the influence of gradient direction while retaining the magnitude information. This allows the method to accurately identify large Gaussians in over-reconstructed areas and split them to recover fine details.
The authors evaluate their proposed AbsGS method on various challenging datasets and demonstrate that it consistently outperforms the original 3D-GS in terms of rendering quality metrics (SSIM, PSNR, LPIPS) while maintaining similar or lower memory consumption. Qualitative results show that AbsGS can effectively eliminate the blurriness caused by over-reconstruction and recover sharp details.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문