교환형 하이브리드 가우시안 기반 고정밀·고품질 3D 재구성
📝 Abstract
Figure 1 : Comparison of 3DGS, 2DGS, and our EGGS. While 3DGS achieves high-fidelity appearance, it often produces inaccurate geometry, with imprecise surfaces and blurred edges. 2DGS improves geometric consistency across views but suffers from reduced appearance quality due to over-smoothed surfaces and loss of detail. In contrast, EGGS employs an exchangeable hybrid Gaussian representation that achieves both accurate geometry and high-quality appearance.
💡 Analysis
Figure 1 : Comparison of 3DGS, 2DGS, and our EGGS. While 3DGS achieves high-fidelity appearance, it often produces inaccurate geometry, with imprecise surfaces and blurred edges. 2DGS improves geometric consistency across views but suffers from reduced appearance quality due to over-smoothed surfaces and loss of detail. In contrast, EGGS employs an exchangeable hybrid Gaussian representation that achieves both accurate geometry and high-quality appearance.
📄 Content
Novel view synthesis (NVS) is a fundamental task in computer graphics and computer vision, with broad applications in augmented reality (AR), virtual reality (VR), and autonomous driving [1,2,3]. Neural Radiance Fields (NeRF) [4] reconstruct implicit radiance fields via differentiable volume rendering. Despite achieving photorealistic appearance and accurate geometry, NeRF-based methods [5,6,7,8,9,10] typically suffer from long training times and slow rendering speeds. 3D Gaussian Splatting (3DGS) [11] has emerged as an efficient alternative, leveraging anisotropic [11] 3D 3D ✗ -General SuGaR CVPR'24 [14] 3D 3D ✗ Normal General GaussianPro ICML'24 [13] 3D 3D ✗ Normal General 2DGS SIGGRAPH'24 [12] 2D 2D ✗ Depth & Normal General GS Surfels SIGGRAPH'24 [15] 2D 3D ⋆ ✗ Depth & Normal General TextureGS ECCV'24 [16] 2D 2D ✗ Depth & Normal General
HybridGS CVPR'25 [17] 3D + 2D* 3D ✗ -Transient HorizonGS CVPR'25 [18] 3D / 2D 3D / 2D ✗ Depth & Normal Varying-altitude
Figure 2: Left: Comparison of 3DGS and 2DGS in appearance and geometry metrics. Right: Comparison between EGGS and related works. Prior works either use only single representation or do not explore complementary advantages of 3D and 2D Gaussians. ⋆ Gaussian Surfel [15] directly sets the z-scale of 3D Gaussian to zero and uses the rasterizer from 3DGS. * HybridGS [17] uses image-frame single-view 2D Gaussians [19,20] instead of 2D Gaussians in the 3D space [12].
3D Gaussians for real-time, high-quality rendering. While 3DGS excels in appearance fidelity, its anisotropic nature often leads to multi-view inconsistencies, limiting geometric accuracy [12,13]. As shown in Figure 1, this can lead to inaccurate edges and surfaces.
Following 3DGS, a line of work has focused on improving its geometric accuracy and reconstruction quality through additional regularization and novel representations, as shown in Figure 2 (right). SUGAR [14] and GaussianPro [13] introduce normal-based regularization, such as planar loss, to align Gaussian normals and encourage flatter shapes, thereby improving surface consistency. Gaussian Surfles [15] and GOF [21] incorporate additional geometry-aware constraints to enhance spatial coherence. 2D Gaussian Splatting (2DGS) [12] replaces 3D ellipsoids with 2D surfels, significantly improving multi-view consistency and geometric accuracy, as shown in Figure 2 (left). However, this comes at the cost of degraded appearance quality, as surfel-based representations struggle to preserve high-frequency details. TextureGS [16] attempts to decouple appearance and geometry within the 2DGS framework, but the single representation still limits overall rendering performance. Recently, HybridGS [17] combines 3DGS with image-space 2D Gaussians to address transient objects, but its radiance field remains fully represented by 3D Gaussians. HorizonGS [18], designed for varying-altitude scenes, decodes 2D Gaussians for surface reconstruction and 3D Gaussians for view synthesis separately via an MLP in ScaffoldGS [22]. While effective in their target domains, these methods do not explore a unified hybrid radiance representation. As a result, the complementary strengths of 2DGS and 3DGS in geometry and appearance remain underutilized.
Effectively combining 3D and 2D Gaussians to jointly improve appearance and geometry is non-trivial, as simply mixing the two representations does not necessarily improve reconstruction quality [18].
To start, the geometric accuracy of 2D Gaussians relies on a ray-splat-intersection-based rasterizer designed to enforce multi-view consistency. Using the projection-based 3DGS rasterizer to render 2D Gaussians can lead to suboptimal geometry [15]. Moreover, Gaussian parameters change significantly during training. For instance, 3D Gaussians may flatten to approximate surfaces, while 2D Gaussians may expand volumetrically to capture thin structures or translucent effects. Fixing the Gaussian type throughout optimization can limit the model’s expressiveness. Finally, relying solely on photometric loss is insufficient to balance geometry and appearance. Additional regularization is required to guide the optimization of hybrid representations. Most importantly, the regularization strategy should account for the distinct characteristics of 3D and 2D Gaussians.
In response to these challenges, we introduce Exchangeable Gaussian Splatting (EGGS), an adaptive hybrid representation that unifies 2D and 3D Gaussian splatting in a single framework. EGGS provides a practical and efficient solution for high-quality novel view synthesis and 3D reconstruction.
Our main contributions are as follows:
• To preserve the complementary strengths of 3D and 2D Gaussians, we develop Hybrid Gaussian Rasterization, a unified rendering framework that supports both projection-based and ray-splat-intersection-based rasterization. We implement this framework with CUDA for efficient optimization, and ensure compatibility with existing 3DGS and 2DGS p
This content is AI-processed based on ArXiv data.