EA-3DGS: Efficient and Adaptive 3D Gaussians with Highly Enhanced Quality for outdoor scenes

TL;DR

EA-3DGS tackles memory and initialization bottlenecks in large-scale outdoor scene reconstruction by integrating an adaptive tetrahedral mesh to guide Gaussian initialization, a contribution-aware pruning scheme paired with a curvature-based densification strategy, and vector-quantized parameter coding. The approach yields high-fidelity, real-time rendering on outdoor scenes, outperforming several NeRF- and 3DGS-based baselines across multiple datasets while dramatically reducing storage via codebooks. Extensive experiments on 13 scenes demonstrate robust performance in low-texture and large-scale outdoor environments, with notable improvements on Mill 19, SCUT-CA, and Tanks & Temples. The work advances practical deployability of 3D Gaussian splatting for outdoor large-scale reconstruction and opens avenues for further mesh-guided initialization and compression refinements.

Abstract

Efficient scene representations are essential for many real-world applications, especially those involving spatial measurement. Although current NeRF-based methods have achieved impressive results in reconstructing building-scale scenes, they still suffer from slow training and inference speeds due to time-consuming stochastic sampling. Recently, 3D Gaussian Splatting (3DGS) has demonstrated excellent performance with its high-quality rendering and real-time speed, especially for objects and small-scale scenes. However, in outdoor scenes, its point-based explicit representation lacks an effective adjustment mechanism, and the millions of Gaussian points required often lead to memory constraints during training. To address these challenges, we propose EA-3DGS, a high-quality real-time rendering method designed for outdoor scenes. First, we introduce a mesh structure to regulate the initialization of Gaussian components by leveraging an adaptive tetrahedral mesh that partitions the grid and initializes Gaussian components on each face, effectively capturing geometric structures in low-texture regions. Second, we propose an efficient Gaussian pruning strategy that evaluates each 3D Gaussian's contribution to the view and prunes accordingly. To retain geometry-critical Gaussian points, we also present a structure-aware densification strategy that densifies Gaussian points in low-curvature regions. Additionally, we employ vector quantization for parameter quantization of Gaussian components, significantly reducing disk space requirements with only a minimal impact on rendering quality. Extensive experiments on 13 scenes, including eight from four public datasets (MatrixCity-Aerial, Mill-19, Tanks \& Temples, WHU) and five self-collected scenes acquired through UAV photogrammetry measurement from SCUT-CA and plateau regions, further demonstrate the superiority of our method.

Figures (8)

• Figure 1: Compared to previous NeRF-based methods (such as Mega-NeRF and Switch-NeRF) and 3DGS-based methods (including 3DGS, C3DGS, Scaffold-GS, and GaMeS), we observe that Mega-NeRF turki2022mega and Switch-NeRF zhenxing2022switch display noticeable blurriness and slow rendering speeds. All comparisons were conducted on an NVIDIA 3090 GPU with 24GB of memory. Relevant parameters are shown in the figure.
• Figure 2: Data acquisition at the SCUT campus using a DJI Matrice 300 RTK UAV.
• Figure 3: The overall pipeline of EA-3DGS. Our proposed method begins by generating an adaptive tetrahedral mesh from the input multi-view images, which serves as the foundation for optimizing 3D Gaussians. During training, EA-3DGS assesses the global importance of each Gaussian distribution based on training observations, enabling the pruning of less significant Gaussians. To mitigate issues arising from excessive pruning, we introduce a structure-aware densification strategy that selectively increases Gaussian density in low-curvature regions. Furthermore, to facilitate real-time rendering in larger scenes, we apply the codebook technique to quantize Gaussian component parameters, effectively reducing memory usage.
• Figure 4: Qualitative comparison of our method with other competitive methods on the Mill 19 and MatrixCity datasets. Red, yellow, and green squares highlight the visual differences for clearer comparison. Our model achieves more refined reconstruction, as evidenced in the Building scene. Our model excels in rendering quality, particularly in low-semantic regions such as grid lines.
• Figure 5: Qualitative comparison of our method with other competitive methods on the SCUT-CA scenes. Red, yellow, and blue squares highlight the visual differences for clearer comparison. Our method demonstrates superior fidelity in handling details, which is particularly evident in the School History Museum scene. Specifically, our method achieves more accurate reconstruction of road lane markings and directional signs.