Research map / 3D vision / Gaussian splatting
HoGS: Unified Near and Far Object Reconstruction via Homogeneous Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HoGS, which represents 3D Gaussian positions and scales in homogeneous coordinates within 3DGS to improve reconstruction of distant objects in unbounded scenes while keeping real-time rendering.
Problem
3DGS uses Cartesian coordinates for Gaussian positions and scales, which makes it hard to optimize and render distant objects in unbounded outdoor scenes.
Approach
Reparametrize each Gaussian's mean and scale as homogeneous coordinates sharing a learnable weight w (initialized from 1/distance), so distant Gaussians can be reached and sized consistently; also disable removal of large world-space Gaussians during pruning.
Key results
- On Tanks&Temples far regions, HoGS reaches 30.77 PSNR vs 29.40 for 3DGS and 30.41 for Zip-NeRF.
- On DL3DV-10K, HoGS achieves 29.93 PSNR and 0.919 SSIM overall vs 28.47 PSNR for 3DGS (50K iterations).
- HoGS trains in about 40-60 minutes on a single RTX 8000 versus roughly three hours for Zip-NeRF on four A6000 GPUs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use homogeneous coordinates for Gaussian positions and scales instead of Cartesian 3DGS coordinates | HoGS / 3DGS · Tanks&Temples (Far) | PSNR: 29.40 -> 30.77 | Helped |
| Use homogeneous coordinates for Gaussian positions and scales instead of Cartesian 3DGS coordinates | HoGS / 3DGS · DL3DV-10K (Far) | PSNR: 36.46 -> 40.55 | Helped |
| Use homogeneous coordinates for Gaussian positions and scales instead of Cartesian 3DGS coordinates | HoGS / 3DGS · Mip-NeRF360 (Indoor) | PSNR: 31.24 -> 31.38 | Mixed |
| Remove homogeneous scaling (ablation) | HoGS · HoGS evaluation scenes | PSNR: 22.60 -> 22.48 | Hurt |
| Remove homogeneous scaling (ablation) | HoGS · HoGS evaluation scenes (far) | PSNR: 33.12 -> 33.06 | Hurt |
| Use the original pruning strategy that removes large world-space Gaussians (ablation) | HoGS · HoGS evaluation scenes | PSNR: 22.60 -> 22.31 | Hurt |
| Use the original pruning strategy that removes large world-space Gaussians (ablation) | HoGS · HoGS evaluation scenes (far) | PSNR: 33.12 -> 31.97 | Hurt |
| Initialize homogeneous weight w with different values rather than 1/d (ablation) | HoGS · HoGS evaluation scenes | PSNR: 22.60; 21.78; 22.38; 22.52; 22.33; 22.16 | No clear effect |
| Increase the 3DGS point count to approximately match HoGS (ablation) | 3DGS · TRAIN and DL3DV scenes 69 and 97 | PSNR: 26.06 -> 25.98 | No clear effect |
Surprising findings
- Increasing the number of 3DGS points to match HoGS does not improve its far-region quality, suggesting the gain is not from point count alone.
- HoGS is largely insensitive to the initialization of the homogeneous weight w across values from 0.01 to 100.
Limitations
- Preserving large world-space Gaussians and densifying distant regions increases point count by 28.3% over 3DGS, raising memory and storage.
- Inherits 3DGS limitations on poorly observed regions, producing artifacts.
- Inherits 3DGS visibility algorithm, which can cause abrupt depth-ordering changes.
Experimental setup
Datasets: Mip-NeRF 360 Tanks&Temples DL3DV-10K
Benchmarks: Mip-NeRF 360 Tanks&Temples DL3DV-10K
Metrics: PSNR SSIM LPIPS training time FPS storage
Baselines: INGP Mip-NeRF 360 Zip-NeRF Scaffold-GS 3DGS
Builds on: 3D Gaussian Splatting Zip-NeRF Scaffold-GS Mip-NeRF 360
Closest papers by meaning
- DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · similarity 0.91)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.90)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.90)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.90)
- UVGS: Reimagining Unstructured 3D Gaussian Splatting using UV Mapping (CVPR 2025 · similarity 0.90)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.