AbsGS: Recovering Fine Details for 3D Gaussian Splatting

AbsGS: Recovering Fine Details for 3D Gaussian Splatting

16 Apr 2024 | Zongxin Ye*, Wenyu Li*, Sidun Liu, Peng Qiao, Yong Dou
The paper addresses the issue of over-reconstruction in 3D Gaussian Splatting (3D-GS), a technique used for novel view synthesis. The original 3D-GS uses an adaptive density control strategy to split or clone Gaussians, but this strategy often fails to effectively split large Gaussians in over-reconstructed regions, leading to blurry rendered images. The authors identify the root cause of this issue as "gradient collision," where the sum of pixel-wise sub-gradients results in a small-scale view-space positional gradient, preventing the split of large Gaussians. To address this, the authors propose a novel method called AbsGS (Absorption Gaussian Splatting). AbsGS introduces the concept of homodirectional view-space positional gradient, which is the sum of the absolute values of pixel-wise sub-gradients. This approach avoids gradient collision by ensuring consistent gradient directions, allowing for the accurate identification and splitting of large Gaussians in over-reconstructed regions. Experiments on various datasets show that AbsGS significantly improves the quality of novel view synthesis, achieving higher PSNR, SSIM, and LPIPS scores while maintaining or reducing memory consumption compared to 3D-GS. The method is also shown to be effective in eliminating over-reconstruction and recovering fine details, as demonstrated through visual comparisons and quantitative evaluations. The authors conclude that AbsGS offers a more efficient and accurate solution to the over-reconstruction problem in 3D-GS.The paper addresses the issue of over-reconstruction in 3D Gaussian Splatting (3D-GS), a technique used for novel view synthesis. The original 3D-GS uses an adaptive density control strategy to split or clone Gaussians, but this strategy often fails to effectively split large Gaussians in over-reconstructed regions, leading to blurry rendered images. The authors identify the root cause of this issue as "gradient collision," where the sum of pixel-wise sub-gradients results in a small-scale view-space positional gradient, preventing the split of large Gaussians. To address this, the authors propose a novel method called AbsGS (Absorption Gaussian Splatting). AbsGS introduces the concept of homodirectional view-space positional gradient, which is the sum of the absolute values of pixel-wise sub-gradients. This approach avoids gradient collision by ensuring consistent gradient directions, allowing for the accurate identification and splitting of large Gaussians in over-reconstructed regions. Experiments on various datasets show that AbsGS significantly improves the quality of novel view synthesis, achieving higher PSNR, SSIM, and LPIPS scores while maintaining or reducing memory consumption compared to 3D-GS. The method is also shown to be effective in eliminating over-reconstruction and recovering fine details, as demonstrated through visual comparisons and quantitative evaluations. The authors conclude that AbsGS offers a more efficient and accurate solution to the over-reconstruction problem in 3D-GS.
Reach us at info@study.space
[slides] AbsGS%3A Recovering Fine Details for 3D Gaussian Splatting | StudySpace