Developing a Model for Improving 3D Gaussian Splatting Performance Based on DBSCAN
Published in IEIE 2024, 2024
This paper proposes a model that improves the performance of 3D Gaussian Splatting (3DGS) while reducing memory usage. The model utilizes DBSCAN clustering to optimize Gaussian representations, leading to more efficient and accurate 3D rendering.
| SSIM | PSNR | Mem | num | |
|---|---|---|---|---|
| M-NeRF360 | 0.759 | 22.22 | 8.6MB | - |
| 3DGS | 0.841 | 23.14 | 411MB | 1.78 |
| Ours-1 | 0.887 | 25.57 | 387.6MB | 1.63 |
| Ours-3 | 0.885 | 25.21 | 384.4MB | 1.62 |
The evaluation metrics for image quality, SSIM and PSNR, show that Ours-1 (DBSCAN applied once) and Ours-3 (DBSCAN applied three times) achieved higher similarity than 3DGS.
Additionally, memory usage was reduced by approximately 10% in both Ours-1 and Ours-3 compared to 3DGS.
3DGS_Diet Dayeon Woo‡1, Eunseo Seo‡2, Chehun Han‡3, Yeonkyung Lee‡4, *Changgyun Jin5 (‡: Equally Contributed)
Download Paper | Download Slides