Research map / 3D vision / Gaussian splatting
EAP-GS: Efficient Augmentation of Pointcloud for 3D Gaussian Splatting in Few-shot Scene Reconstruction
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 EAP-GS, which augments SfM point clouds with attention-guided two-view tracks to improve 3DGS initialization for few-shot scene reconstruction.
Problem
3D Gaussian splatting degrades severely with few training views, and the paper argues that inadequate SfM point-cloud initialization (multi-view triangulation discards many points) is a key overlooked cause.
Approach
Attentional Pointcloud Augmentation retains two-view feature tracks only in attention regions (texture-complex but sparsely populated, found via DBSCAN on the existing point density) to add points, then merges them with the standard point cloud for 3DGS initialization; it is modular and can be plugged into other 3DGS few-shot methods.
Key results
- On LLFF (3 views), EAP-GS reaches 18.93 PSNR / 0.6399 SSIM / 0.1792 LPIPS vs. FSGS 19.10 / 0.6246 / 0.1888 and CoR-GS 18.73 / 0.6317 / 0.2214, with 1.75 min training time.
- On Mip-NeRF360 (12 views), EAP-GS reaches 18.08 PSNR / 0.4998 SSIM / 0.3696 LPIPS, versus CoR-GS 17.93 / 0.4892 / 0.4639.
- Adding APA to existing methods improves them, e.g., FSGS on LLFF rises from 19.10 to 19.76 PSNR and CoR-GS from 18.73 to 19.32 PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add APA to 3DGS | 3DGS · LLFF | PSNR: 14.63 -> 14.78 | Helped |
| add APA to 3DGS | 3DGS · Mip-NeRF360 | PSNR: 16.06 -> 17.06 | Helped |
| add APA to DRGS | DRGS · LLFF | PSNR: 17.48 -> 18.12 | Helped |
| add APA to DRGS | DRGS · Mip-NeRF360 | PSNR: 17.11 -> 17.43 | Helped |
| add APA to FSGS | FSGS · LLFF | PSNR: 19.10 -> 19.76 | Helped |
| add APA to FSGS | FSGS · Mip-NeRF360 | PSNR: 17.93 -> 18.51 | Helped |
| add APA to CoR-GS | CoR-GS · LLFF | PSNR: 18.73 -> 19.32 | Helped |
| add APA to CoR-GS | CoR-GS · Mip-NeRF360 | PSNR: 17.93 -> 18.76 | Helped |
| add the attention mechanism to pointcloud augmentation (ablation) | EAP-GS · LLFF | PSNR: 18.78 -> 18.93 | Helped |
| add the attention mechanism to pointcloud augmentation (ablation) | EAP-GS · LLFF | SSIM: 0.6353 -> 0.6399 | Helped |
| add pointcloud augmentation (ablation) | EAP-GS · LLFF | PSNR: 17.43 -> 18.78 | Helped |
| use DetectorfreeSfM rather than COLMAP without APA (ablation) | EAP-GS initialization · LLFF | PSNR: 18.05 -> 17.43 | Hurt |
Surprising findings
- On LLFF, EAP-GS is nearly tied with FSGS on PSNR (18.93 vs 19.10), so it does not lead every metric in the main table despite the claim of leading scores across all metrics.
Limitations
- Improvement in reconstruction quality is limited for extrapolation test views due to data incompleteness.
- Two-view tracks can introduce subtle triangulation errors in special cases, and no mechanism is provided to bound this error.
- Relies on camera poses assumed known; unknown-pose results are reported only briefly and not in the main paper.
Experimental setup
Datasets: LLFF Mip-NeRF360
Benchmarks: LLFF Mip-NeRF360
Metrics: PSNR SSIM LPIPS training time number of Gaussians
Baselines: 3DGS DRGS FSGS CoR-GS
Builds on: 3D Gaussian splatting COLMAP DetectorfreeSfM FSGS DRGS
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.90)
- ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.90)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.90)
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.