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

Jianlin Guo, Haihong Xiao, Wenxiong KangarXiv 2025 · DOI · arXiv

Introduction

Reconstructing large-scale outdoor scenes—such as city blocks or expansive terrains—has long been a significant challenge in computer vision. While Neural Radiance Fields (NeRFs) produce high-quality results, their slow rendering speeds limit practical applications. 3D Gaussian Splatting (3DGS) emerged as a faster, explicit alternative; however, it faces its own hurdles: substantial VRAM and disk space consumption, and difficulty capturing geometry in low-texture areas like flat walls or clear skies.

In the paper "EA-3DGS: Efficient and Adaptive 3D Gaussians with Highly Enhanced Quality for outdoor scenes," researchers from the South China University of Technology introduce a framework designed to make 3DGS both leaner and more geometrically robust for large-scale environments.

The Challenges of Outdoor 3DGS

Standard 3DGS relies on sparse point clouds from Structure-from-Motion (SfM) for initialization. In outdoor settings, these point clouds often contain gaps in regions with repetitive or weak textures. Without high-quality initial points, 3DGS tends to produce blurry or "needle-like" artifacts. Additionally, outdoor scenes require millions of Gaussians, creating two major bottlenecks:

  1. Memory (VRAM): Training can exceed the capacity of standard GPUs.
  2. Storage: High-fidelity scenes can occupy gigabytes of disk space, hindering real-time web or mobile deployment.

Methodology

The authors propose three key innovations to address these limitations:

Overview of the EA-3DGS pipeline, integrating adaptive mesh initialization, contribution-aware pruning, and vector quantization.

1. Adaptive Tetrahedral Mesh Initialization

Rather than initializing Gaussians solely on sparse SfM points, EA-3DGS employs Delaunay triangulation to construct a tetrahedral mesh. By "binding" Gaussians to the faces of this mesh, the system creates a geometric scaffold. This ensures consistent Gaussian placement even in low-texture regions where SfM typically fails to generate points.

2. Dual Optimization: Pruning & Densification

To maintain efficiency, the framework uses a "contribution-aware" pruning strategy. A GS Score is calculated for each Gaussian based on its actual contribution to the final rendered pixels across all training views. Gaussians with low scores are removed.

Conversely, the authors introduce Structure-aware Densification. By calculating local scene curvature, the system strategically adds Gaussians to low-curvature (flat) regions. These areas are often under-represented in standard 3DGS, and this targeted densification helps sharpen fine details.

3. Vector Quantization (VQ)

To address storage concerns, EA-3DGS applies Vector Quantization. Instead of storing unique parameters (color, rotation, scale) for every Gaussian, the model utilizes a "codebook" of representative values. Each Gaussian then stores an index to this codebook, effectively compressing the 3D scene representation.

Datasets and Evaluation

The researchers evaluated EA-3DGS across a wide range of datasets:

  • Public Benchmarks: Mill 19 (Rubble and Building), MatrixCity, WHU (large-scale aerial), and Tanks & Temples.
  • Self-Collected Data: Five diverse scenes captured via DJI Matrice 300 RTK UAVs, including the SCUT campus and high-altitude plateau regions in Tibet, featuring challenging repetitive textures and complex architecture.

Key Results

Visual comparison of the scene before and after 5.2x compression, showing minimal degradation in PSNR and visual quality.
  • Enhanced Quality: On the Mill 19 "Building" scene, EA-3DGS achieved a PSNR of 21.58, outperforming vanilla 3DGS (20.11) and Mega-NeRF (20.93).
  • Significant Compression: Vector quantization reduced a scene's size from 1.8GB to 345MB (a 5.2x compression ratio) with negligible loss in visual quality.
  • Training Efficiency: The pruning strategy enables the optimization of large scenes on a single NVIDIA 3090 GPU (24GB VRAM), a task often impossible for standard 3DGS at this scale.

Conclusion

EA-3DGS offers a robust solution for the limitations of 3D Gaussian Splatting in large-scale outdoor environments. By anchoring Gaussians to an adaptive mesh and employing intelligent pruning and compression, it achieves high-fidelity reconstructions that are both memory-efficient and storage-friendly. While the method remains dependent on the quality of initial SfM data, it represents a significant advancement toward practical, real-time 3D reconstruction of complex real-world scenes.

Summary updated

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.
BibTeX
@misc{guo2025ea,
  author = {Jianlin Guo and Haihong Xiao and Wenxiong Kang},
  title = {{EA-3DGS: Efficient and Adaptive 3D Gaussians with Highly Enhanced Quality for outdoor scenes}},
  year = {2025},
  eprint = {2505.10787},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.10787},
}
Figure