Pixel-GS: Density Control with Pixel-aware Gradient for 3D Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) has demonstrated impressive novel view synthesis results while advancing real-time rendering performance. However, it relies heavily on the quality of the initial point cloud, resulting in blurring and needle-like artifacts in areas with insufficient initializing points. This is mainly attributed to the point cloud growth condition in 3DGS that only considers the average gradient magnitude of points from observable views, thereby failing to grow for large Gaussians that are observable for many viewpoints while many of them are only covered in the boundaries. To this end, we propose a novel method, named Pixel-GS, to take into account the number of pixels covered by the Gaussian in each view during the computation of the growth condition. We regard the covered pixel numbers as the weights to dynamically average the gradients from different views, such that the growth of large Gaussians can be prompted. As a result, points within the areas with insufficient initializing points can be grown more effectively, leading to a more accurate and detailed reconstruction. In addition, we propose a simple yet effective strategy to scale the gradient field according to the distance to the camera, to suppress the growth of floaters near the camera. Extensive experiments both qualitatively and quantitatively demonstrate that our method achieves state-of-the-art rendering quality while maintaining real-time rendering speed, on the challenging Mip-NeRF 360 and Tanks & Temples datasets.

Figures (6)

• Figure 1: Our Pixel-GS effectively grows points in areas with insufficient initializing points (a), leading to a more accurate and detailed reconstruction (d). In contrast, 3D Gaussian Splatting (3DGS) suffers from blurring and needle-like artifacts in these areas, even with a lower threshold of splitting and cloning to encourage more grown points (c). The rendering quality (in LPIPS $\downarrow$) and memory consumption are shown in the results. 3DGS$^*$ is our retrained 3DGS model with better performance.
• Figure 2: Pipeline of Pixel-GS.$\mathrm{p}_i$ represents the number of pixels participating in the calculation for the Gaussian from this viewpoint, and $\mathbf{g}_i$ represents the gradient of the Gaussian's NDC coordinates. We changed the condition for deciding whether a Gaussian should split or clone from the left to the right side.
• Figure 3: Qualitative comparison between Pixel-GS (Ours) and 3DGS$^*$. The first three scenes are from the Mip-NeRF 360 dataset (Bicycle, Flowers, and Treehill), while the last four scenes are from the Tanks & Temples dataset (Barn, Caterpillar, Playground, and Train). The blow-up regions or arrows highlight the parts with distinct differences in quality. 3DGS$^*$ is our retrained 3DGS model with better performance.
• Figure 4: Qualitative results of the ablation study. The PSNR$\uparrow$ results are shown on the corresponding images.
• Figure 5: Reconstruction quality (PSNR$\uparrow$, SSIM$\uparrow$, and LPIPS$\downarrow$) vs. Dropping rate of initializing points. Here, the dropping rate refers to the percentage of points dropped from the original SfM point clouds for initializing Gaussians. The results are obtained on the Mip-NeRF 360 dataset.