Research map / 3D vision / Gaussian splatting
Decoupling Appearance Variations with 3D Consistent Features in 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 DAVIGS, a decoupled image-level appearance module for Gaussian Splatting that predicts per-pixel affine color transforms from hash-grid 3D features and per-view embeddings, with cell-based querying and an identity regularizer.
Problem
Appearance variations (camera ISP, time of day, weather, local lighting) cause floaters and color distortions in Gaussian Splatting reconstructions; existing decoupled methods handle only mild global changes while coupled methods are slow and memory-heavy.
Approach
Transform rendered images at the image level (decoupled, discardable after optimization) using per-pixel affine matrices decoded from multi-resolution hash-grid features of back-projected 3D coordinates plus a per-view appearance embedding, giving 3D-consistent local appearance handling; a cell-based query and identity regularization make it efficient and stable.
Key results
- On the GLAV dataset DAVIGS reaches 28.31 PSNR / 0.861 SSIM / 0.134 LPIPS with 19 min training, 3.4G VRAM and 151 FPS, outperforming GS-W (27.23 PSNR) and WildGaussians (24.68 PSNR).
- As a plug-and-play module, it raises 2DGS PSNR from 22.70 to 27.76 and GOF PSNR from 22.38 to 28.21 with small extra training time.
- Replacing 3D xyz features with uv, depth, or color features drops PSNR to 20.04-22.12 from 22.75 in the ablation (ablation reported on a different scale than main table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace 3D xyz features with uv coordinates (ablation) | DAVIGS · GLAV | PSNR: 22.75 -> 21.76 | Hurt |
| replace 3D xyz features with depth (ablation) | DAVIGS · GLAV | PSNR: 22.75 -> 20.04 | Hurt |
| replace 3D xyz features with uv coordinates and depth (ablation) | DAVIGS · GLAV | PSNR: 22.75 -> 22.12 | Hurt |
| replace 3D xyz features with RGB color (ablation) | DAVIGS · GLAV | PSNR: 22.75 -> 21.58 | Hurt |
| add the transformation regularizer L_ID (ablation) | DAVIGS · GLAV | PSNR: 21.33 -> 22.75 | Helped |
| set query cell size to 1 instead of 8 (ablation) | DAVIGS · GLAV | PSNR: 28.31 -> 28.24 | No clear effect |
| set query cell size to 2 instead of 8 (ablation) | DAVIGS · GLAV | PSNR: 28.31 -> 28.35 | No clear effect |
| set query cell size to 4 instead of 8 (ablation) | DAVIGS · GLAV | PSNR: 28.31 -> 28.33 | No clear effect |
| set query cell size to 16 instead of 8 (ablation) | DAVIGS · GLAV | PSNR: 28.31 -> 28.18 | Hurt |
| set query cell size to 32 instead of 8 (ablation) | DAVIGS · GLAV | PSNR: 28.31 -> 27.76 | Hurt |
Surprising findings
- Plugging the module into baselines reduced VRAM in some cases (e.g., GOF from 18.11G to 12.43G), attributed to suppressed floaters.
Limitations
- Scaling to larger scenes may require larger hash grids and deeper MLPs, increasing VRAM, storage and optimization time.
- Evaluated mainly on a self-collected dataset of seven scenes plus PhotoTourism.
Experimental setup
Datasets: GLAV PhotoTourism Mill-19
Benchmarks: GLAV PhotoTourism
Metrics: PSNR SSIM LPIPS training time VRAM FPS
Baselines: NeRF-W Ha-NeRF 3DGS GS-W WildGaussians VastGaussian Mip-Splatting 2DGS GOF
Builds on: 3D Gaussian Splatting multi-resolution hash encoding VastGaussian Robust Gaussian Splatting Mip-Splatting
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.91)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.91)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (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.