Research map / 3D vision / Gaussian splatting
Perceptual-GS: Scene-adaptive Perceptual Densification for 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.
Perceptual-GS guides 3DGS densification with human-perception-based sensitivity maps (Sobel edges, binarized and smoothed) learned per Gaussian through a dual-branch rendering, producing higher perceptual quality with fewer Gaussians.
Problem
3DGS's adaptive density control distributes Gaussians poorly, causing blur in detailed regions or redundancy in simple ones, and existing densification refinements struggle to balance quality and efficiency, especially in large-scale scenes.
Approach
Learn a per-Gaussian perceptual sensitivity parameter via a sensitivity branch that renders binarized, perception-enhanced gradient maps; use it to densify high- and medium-sensitivity Gaussians, reinitialize depth for sparse-SfM scenes, and reduce clone-induced redundancy via an opacity decline transform.
Key results
- On Mip-NeRF 360, Perceptual-GS reaches 28.01 PSNR / 0.839 SSIM / 0.172 LPIPS with 2.69M Gaussians vs. 3DGS* 27.71 / 0.826 / 0.202 with 3.14M.
- On BungeeNeRF, Perceptual-GS reaches 27.86 PSNR / 0.918 SSIM / 0.095 LPIPS with 4.97M Gaussians vs. 3DGS* 27.64 / 0.912 / 0.100 with 6.92M, while Pixel-GS and Taming-3DGS hit OOM in some scenes.
- Integrated with Pixel-GS on Mip-NeRF 360, LPIPS drops from 0.176 to 0.167 while Gaussians fall from 5.23M to 3.37M.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply Perceptual-GS densification and representation to 3DGS | 3DGS · Mip-NeRF 360 | LPIPS: 0.202 -> 0.172 | Helped |
| Apply Perceptual-GS densification and representation to 3DGS | 3DGS · Tanks & Temples | LPIPS: 0.178 -> 0.151 | Helped |
| Apply Perceptual-GS densification and representation to 3DGS | 3DGS · Deep Blending | LPIPS: 0.247 -> 0.231 | Helped |
| Apply Perceptual-GS densification and representation to 3DGS | 3DGS · BungeeNeRF | LPIPS: 0.100 -> 0.095 | Helped |
| Add perception-oriented enhancement to sensitivity maps (ablation) | Perceptual-GS · Mip-NeRF 360 | PSNR: 27.74 -> 28.01 | Helped |
| Add high-sensitivity Gaussian densification (ablation) | Perceptual-GS · Mip-NeRF 360 | PSNR: 27.74 -> 28.01 | Helped |
| Add medium-sensitivity Gaussian densification (ablation) | Perceptual-GS · Mip-NeRF 360 | PSNR: 27.86 -> 28.01 | Helped |
| Add scene-adaptive depth reinitialization (ablation) | Perceptual-GS · Mip-NeRF 360 | PSNR: 27.93 -> 28.01 | Helped |
| Add opacity decline to clone operation (ablation) | Perceptual-GS · Mip-NeRF 360 | #G: 3.25M -> 2.69M | Helped |
| Increase opacity decline exponent to k=2.0 (ablation) | Perceptual-GS · Mip-NeRF 360 | LPIPS: 0.172 -> 0.174 | Hurt |
| Increase sensitivity loss weight λS from 0.1 to 0.3 (ablation) | Perceptual-GS · Mip-NeRF 360 | PSNR: 28.01 -> 27.82 | Hurt |
| Increase sensitivity loss weight λS from 0.1 to 0.5 (ablation) | Perceptual-GS · Mip-NeRF 360 | PSNR: 28.01 -> 27.48 | Hurt |
Surprising findings
- Dual-branch rendering alone reduces Gaussian count, since low-sensitivity Gaussians get lower sensitivity loss and fall below the position-gradient densification threshold.
- Perceptual-GS does not improve PSNR on some datasets (e.g., Deep Blending PSNR 29.94 vs Taming-3DGS 30.14) and the w/o OD variant preserves quality at higher Gaussian count.
Limitations
- Relies on hand-designed binarization and smoothing thresholds (tau_e, tau_s) and many tuned hyperparameters.
- Per-scene PSNR gains on some scenes are negative (e.g., Flowers, Treehill), and some metrics are weaker than baselines on certain datasets.
Experimental setup
Datasets: Mip-NeRF 360 Tanks & Temples Deep Blending BungeeNeRF
Benchmarks: Mip-NeRF 360 Tanks & Temples Deep Blending BungeeNeRF
Metrics: PSNR SSIM LPIPS #Gaussians FPS QEB
Baselines: 3DGS Pixel-GS Mini-Splatting-D Taming-3DGS
Builds on: 3D Gaussian Splatting Pixel-GS Mini-Splatting-D CoR-GS Taming-3DGS
Closest papers by meaning
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.91)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.91)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- 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.