Research map / Point clouds / Point cloud classification
Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians
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.
Predicts 2D Gaussians directly from point clouds with normal-based initialization, splitting decoders, and an entire-patch architecture, enabling sparse-point and cross-category photo-realistic point cloud rendering without image refinement.
Problem
Photo-realistic rendering of sparse point clouds without per-scene training, while generalizing across categories and point densities; prior learning-based methods depend on categorical priors, dense points, or image refinement.
Approach
Predict 2D Gaussians (with normals initialized from estimated point normals) from point clouds; split each point into K Gaussians via splitting decoders to densify sparse inputs; train on point-cloud patches using an entire-cloud module that supplies background Gaussians so rendered supervision is complete.
Key results
- Achieves average improvements of 2.60% PSNR, 1.05% SSIM and 6.06% LPIPS over baselines across datasets (e.g., Car PSNR 28.73 at 100K vs. PFGS 27.34).
- Renders at 142.86 FPS on RTX 3090, versus 37.45 for NPBG++, 3.80 for PFGS and 1.62 for TriVol.
- Using only 20K points, the method still beats baselines on chairs, tables and shoes; ablation without initialization drops PSNR from 27.88 to 9.62.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed 2D Gaussian point-cloud rendering method instead of PFGS | Point-cloud rendering · Car (ShapeNet) | PSNR: 27.34 -> 28.73 | Helped |
| Use the proposed 2D Gaussian point-cloud rendering method instead of TriVol | Point-cloud rendering · Chair (ShapeNet) | PSNR: 28.85 -> 29.10 | Helped |
| Use the proposed 2D Gaussian point-cloud rendering method instead of TriVol | Point-cloud rendering · Shoe (GSO) | PSNR: 31.24 -> 32.08 | Helped |
| Use the proposed 2D Gaussian point-cloud rendering method instead of PFGS | Point-cloud rendering · ScanNet | LPIPS: 0.452 -> 0.490 | Hurt |
| Add point-cloud-informed Gaussian initialization (ablation) | 2D Gaussian prediction module · Car (ShapeNet) | PSNR: 9.62 -> 27.88 | Helped |
| Add entire-cloud module Ne alongside the patch module (ablation) | Entire-patch architecture · Car (ShapeNet) | PSNR: 12.94 -> 27.88 | Helped |
| Add entire-cloud module Ne rather than using Np only (ablation) | Entire-patch architecture · Car (ShapeNet) | PSNR: 23.23 -> 27.88 | Helped |
| Increase the Gaussian split number K (ablation) | Splitting decoder · Car (ShapeNet) | PSNR | Helped |
Surprising findings
- Trained only on the Car category, the model generalizes to DTU at 16.7 PSNR, where PFGS trained on Car with 20K points reaches only 8.07 PSNR.
Limitations
- Cannot render a complete image when part of the point cloud is absent.
- Not yet tested on noisy point clouds scanned in the wild or outdoor scenes.
Experimental setup
Datasets: ShapeNet GSO ScanNet DTU THuman2.0
Benchmarks: ScanNet ShapeNet Car ShapeNet Chair GSO Shoe THuman2.0 DTU
Metrics: PSNR SSIM LPIPS FPS
Baselines: NPBG++ TriVol PFGS Traditional graphics rendering
Builds on: 2D Gaussian Splatting PointMLP PFGS TriVol
Closest papers by meaning
- GaussianPainter: Painting Point Cloud into 3D Gaussians with Normal Guidance (AAAI 2025 · similarity 0.90)
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (CVPR 2025 · similarity 0.89)
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.89)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.88)
- LeanGaussian: Breaking Pixel or Point Cloud Correspondence in Modeling 3D Gaussians (CVPR 2025 · similarity 0.88)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.87)
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.