Research map / 3D vision / Gaussian splatting
Towards Realistic Example-based Modeling via 3D Gaussian Stitching
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 an interactive GUI-driven method that stitches multiple 3D Gaussian Splatting scenes into realistic composites using KNN boundary detection and a two-phase sampling-based cloning plus clustering-based color tuning optimization.
Problem
Realistic example-based modeling of real-world captured 3D objects requires seamless appearance blending when composing parts, which prior grid-based NeRF stitching (SeamlessNeRF) cannot do interactively or on real-world scenes.
Approach
Replace grid-based gradient propagation with point-based 3D Gaussian Splatting: identify boundary Gaussians via KNN against the source field, propagate color/SH features to remaining target points by sampling driven points, add a Sobel-based 2D gradient constraint, and apply a palette-based global color tuning phase, all within a real-time GUI.
Key results
- VQA average score of 0.836 for the proposed method versus 0.662 for SeamlessNeRF (Table 1).
- SeamlessNeRF reported to fail in all three real-world comparison scenes, with artifacts and weak color propagation.
- Ablations show the S-phase and T-phase are complementary: T-phase corrects global brightness/hue/saturation left by S-phase alone; the full pipeline converges in about 4 minutes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 3D Gaussian Stitching instead of SeamlessNeRF | Real-world Gaussian stitching results · VQA evaluation | VQA average score: 0.662 -> 0.836 | Helped |
| add clustering-based tuning (T-phase) to sampling-based cloning (S-phase) (ablation) | 3D Gaussian Stitching · Gaussian stitching | global appearance harmony | Helped |
| use sampling-based cloning (S-phase) together with clustering-based tuning (T-phase), rather than either phase alone (ablation) | 3D Gaussian Stitching · Gaussian stitching | stitching quality | Helped |
| remove color loss from the S-phase (ablation) | 3D Gaussian Stitching · Gaussian stitching | feature propagation | Hurt |
| increase the gradient-loss weight λ1 (ablation) | 3D Gaussian Stitching · Gaussian stitching | content preservation and harmonization: λ1=0.1, λ1=2, λ1=10 | Mixed |
| use 2D screen-space gradient supervision with the Sobel operator instead of straightforward gradient loss (ablation) | 3D Gaussian Stitching · Gaussian stitching | gradient-loss effectiveness | Helped |
| remove the mapping function φ from the S-phase (ablation) | 3D Gaussian Stitching · Gaussian stitching | composition realism | Hurt |
| remove the sampling strategy for propagating view-dependent effects in the S-phase (ablation) | 3D Gaussian Stitching · Gaussian stitching | view-dependent effect propagation | Hurt |
Surprising findings
- Naive Laplacian-coordinate regularization fails even before propagation begins, motivating the sampling-based strategy.
- Standard 2D style transfer (SNeRF-style, re-implemented on 3DGS) fails to produce seamless stitching.
Limitations
- Only rigid transformations; no non-rigid deformation (ARAP suggested as future work).
- Does not handle specular reflections or translucent objects; no PBR decoupling.
- Quantitative evaluation is weak: uses a VQA-based score with no ground truth or specialized realism metric.
Experimental setup
Datasets: BlendedMVS Mip360 SeamlessNeRF dataset
Benchmarks: BlendedMVS Mip360 SeamlessNeRF dataset
Metrics: VQA average score
Baselines: SeamlessNeRF SNeRF-style 2D style transfer
Builds on: 3D Gaussian Splatting SeamlessNeRF Poisson-based seamless editing Sobel operator
Closest papers by meaning
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.89)
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.88)
- Multi-StyleGS: Stylized Gaussian Splatting with Multiple Styles (AAAI 2025 · similarity 0.88)
- Hierarchical Gaussian Mixture Model Splatting for Efficient and Part Controllable 3D Generation (CVPR 2025 · similarity 0.88)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · 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.