HoGS: Unified Near and Far Object Reconstruction via Homogeneous Gaussian Splatting

Xinpeng Liu, Zeyi Huang, Fumio Okura, Yasuyuki Matsushita

Introduction

Reconstructing large-scale outdoor environments in 3D remains a significant challenge for computer vision. While 3D Gaussian Splatting (3DGS) has set new standards for real-time, photorealistic rendering, it often struggles with "unbounded" scenes where objects range from immediate foreground to distant horizons. This paper introduces HoGS (Homogeneous Gaussian Splatting), a method that leverages projective geometry to unify the representation of near and far objects.

The Distance Problem in 3DGS

Standard 3DGS defines objects using Cartesian coordinates $(x, y, z)$. In outdoor settings, distant elements like mountains or clouds occupy very small areas on the image plane due to perspective projection. This makes it mathematically difficult for optimization algorithms to accurately place or scale Gaussians in far-off 3D space. Consequently, backgrounds in standard 3DGS often appear as blurry, indistinct blobs or are ignored entirely.

HoGS (Ours) demonstrates sharp reconstruction of both the foreground train and the distant background hills (orange box).

Methodology: The Power of Homogeneous Coordinates

The core innovation of HoGS is the transition from Cartesian to Projective Geometry.

  1. Homogeneous Position: HoGS represents the mean position of each Gaussian as $(\tilde{x}, \tilde{y}, \tilde{z}, w)$. As the weight $w$ approaches zero, the point moves toward infinity. This allows the model to represent points at any distance within a finite coordinate space.
  2. Homogeneous Scaling: Crucially, the authors apply this logic to the scale of the Gaussians. By tying the scale to the $w$ component, a Gaussian's size naturally scales with its distance, ensuring it remains observable on the image plane during optimization.
  3. Adaptive Pruning: Unlike standard 3DGS, which deletes large Gaussians to reduce clutter, HoGS allows large-diameter Gaussians to persist if they are positioned far away, recognizing their necessity for representing distant backgrounds.

Evaluation and Results

HoGS was evaluated on benchmarks including Mip-NeRF 360, Tanks & Temples, and DL3DV-10K. The results demonstrate a clear advantage in capturing far-field details.

Standard 3DGS struggles with distant objects, resulting in a lower PSNR and a noticeably blurrier background compared to HoGS.

On the DL3DV "Far" benchmark, HoGS achieved a PSNR of 40.55, significantly outperforming standard 3DGS (36.46). Despite the increased complexity of the scene representation, HoGS maintains the real-time performance characteristic of splatting, consistently achieving over 200 FPS. Training is also efficient, typically requiring only 40–60 minutes on a single GPU—a fraction of the time needed for high-quality NeRF alternatives like Zip-NeRF.

Conclusion

By aligning the underlying 3D representation with the projective nature of cameras, HoGS provides a simple yet powerful solution for unbounded scene reconstruction. It eliminates the need for complex pre-processing like skybox segmentation, offering a unified, high-performance framework for both near and far object reconstruction.

Summary updated

Abstract

Novel view synthesis has demonstrated impressive progress recently, with 3D Gaussian splatting (3DGS) offering efficient training time and photorealistic real-time rendering. However, reliance on Cartesian coordinates limits 3DGS's performance on distant objects, which is important for reconstructing unbounded outdoor environments. We found that, despite its ultimate simplicity, using homogeneous coordinates, a concept on the projective geometry, for the 3DGS pipeline remarkably improves the rendering accuracies of distant objects. We therefore propose Homogeneous Gaussian Splatting (HoGS) incorporating homogeneous coordinates into the 3DGS framework, providing a unified representation for enhancing near and distant objects. HoGS effectively manages both expansive spatial positions and scales particularly in outdoor unbounded environments by adopting projective geometry principles. Experiments show that HoGS significantly enhances accuracy in reconstructing distant objects while maintaining high-quality rendering of nearby objects, along with fast training speed and real-time rendering capability. Our implementations are available on our project page https://kh129.github.io/hogs/.
BibTeX
@inproceedings{liu2025hogs,
  author = {Xinpeng Liu and Zeyi Huang and Fumio Okura and Yasuyuki Matsushita},
  title = {{HoGS: Unified Near and Far Object Reconstruction via Homogeneous Gaussian Splatting}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.02488},
}
Figure