Gaussian Billboards: Expressive 2D Gaussian Splatting with Textures

TL;DR

This work introduces Gaussian Billboards, an enhancement to 2D Gaussian Splatting that replaces each splat's solid color with a small per-primitive texture grid defined over a limited uv extent. By performing bilinear color interpolation tied to the splat's uv- coordinates, the method expresses richer surface textures without increasing the primitive count, improving image fidelity and 3D reconstruction quality. The authors provide detailed ablations on texture extent and resolution, demonstrate gains on image fitting, and show improved 3D reconstructions on NeRF360 and face datasets, with practical considerations for training efficiency. The approach is positioned as orthogonal to other refinements and may synergize with methods that modulate opacity or texture globally.

Abstract

Gaussian Splatting has recently emerged as the go-to representation for reconstructing and rendering 3D scenes. The transition from 3D to 2D Gaussian primitives has further improved multi-view consistency and surface reconstruction accuracy. In this work we highlight the similarity between 2D Gaussian Splatting (2DGS) and billboards from traditional computer graphics. Both use flat semi-transparent 2D geometry that is positioned, oriented and scaled in 3D space. However 2DGS uses a solid color per splat and an opacity modulated by a Gaussian distribution, where billboards are more expressive, modulating the color with a uv-parameterized texture. We propose to unify these concepts by presenting Gaussian Billboards, a modification of 2DGS to add spatially-varying color achieved using per-splat texture interpolation. The result is a mixture of the two representations, which benefits from both the robust scene optimization power of 2DGS and the expressiveness of texture mapping. We show that our method can improve the sharpness and quality of the scene representation in a wide range of qualitative and quantitative evaluations compared to the original 2DGS implementation.

Figures (8)

• Figure 1: We propose an enhancement for 2D Gaussian Splatting that (a) replaces the per-primitive solid color with a small per-primitive texture. This allows the 2DGS to represent more details (b) when the number of primitives is fixed to, e.g., 100000. Even with the standard densification strategy and identical hyperparameters, it improves the quality of 3D scene reconstruction (c), shown on a test view.
• Figure 2: The color grid of Gaussian Billboards spans the $uv$-parametrization of the 2D primitive in the range of $[-\sigma,+\sigma]$ with $\sigma=0.5$. This way, the border of the grid where the interpolation starts being clamped roughly corresponds to the isoline of $0.8$ opacity. On the right you can see how the texture within a primitive is rotated and scaled with $s_u,s_v,\mathbf{t}_u,\mathbf{t}_v$.
• Figure 3: The three images used for image fitting tests and hyperparameter ablations: Creation of Adam, Meteora, and Zurich. The marked areas indicate the crops visualized in \ref{['fig:image-texture-resolution']}.
• Figure 4: Overfitting a single image (a) with a fixed number of 1000 splats. For that limited number of splats, Ours with $N=4$ (c) achieves better quality than Traditional 2DGS (b) in terms of quantitative statistics and image sharpness. If we assign random colors, the individual splats and the bilinear interpolation of the color grid become visible (d,e). To verify that the optimization utilizes the additional color information, we downsample the color grid to a resolution of $2 \times 2$ (f) and a single color (g).
• Figure 5: Ablation of the spatial extent of the grid in the $uv$-plane, $\sigma$, for overfitting of Creation of Adam and Meteora (see \ref{['fig:results:input-images']}) with $N=4$. With a lower value for $\sigma$, the texture information is concentrated in the area of the primitive with the highest opacity. As can be seen in the plot, the optimal value for $\sigma$ is achieved at around $0.5$. For a visualization of what different values for $\sigma$ mean, the bottom row shows three Gaussian primitives with a random texture rendered with four values for $\sigma$.