Research map / 3D vision / Gaussian splatting
Wild-GS: Real-Time Novel View Synthesis from Unconstrained Photo Collections
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.
Wild-GS adapts 3D Gaussian Splatting to unconstrained photo collections by combining global embeddings, triplane-based local appearance features sampled per Gaussian, learnable intrinsic features, a transient visibility mask, and depth regularization.
Problem
Novel view synthesis from in-the-wild tourist photo collections, which have variable appearance and transient occluders; prior NeRF-based methods are slow to train and render.
Approach
Align reference-image appearance to 3D Gaussians explicitly by back-projecting the static reference view into a point cloud, encoding it into triplane features, and having each Gaussian sample its local appearance from the triplane; the fused global, local, and intrinsic features decode SH coefficients, so rasterization stays unchanged and renders stay fast.
Key results
- Wild-GS reaches 29.65 PSNR on Brandenburg Gate, 24.99 on Sacre Coeur, and 24.45 on Trevi Fountain, the best PSNR and SSIM among compared methods.
- Wild-GS trains in about 0.52 hours at 227 FPS inference on a single RTX3090, versus 101 hours and 0.02 FPS for CR-NeRF per the table's efficiency columns.
- Cropping the triplane with AABB cuts training time by about 39% while keeping or improving rendering quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add triplane cropping (ablation) | Wild-GS · Phototourism | PSNR: 29.15 -> 29.65; 24.93 -> 24.99; 23.61 -> 24.45 | Helped |
| Add triplane cropping (ablation) | Wild-GS · Phototourism | Training time (hours): 0.85 -> 0.52 | Helped |
| Add reverse-direction triplane projections (ablation) | Wild-GS · Phototourism | PSNR: 28.52 -> 29.65; 24.37 -> 24.99; 24.14 -> 24.45 | Helped |
| Add global appearance encoding (ablation) | Wild-GS · Phototourism | PSNR: 29.00 -> 29.65; 24.82 -> 24.99; 24.18 -> 24.45 | Helped |
| Add transient mask prediction (ablation) | Wild-GS · Phototourism | PSNR: 28.85 -> 29.65; 24.71 -> 24.99; 23.91 -> 24.45 | Helped |
| Add depth regularization (ablation) | Wild-GS · Phototourism | PSNR: 28.36 -> 29.65; 23.36 -> 24.99; 23.20 -> 24.45 | Helped |
Surprising findings
- Global appearance encoding gives no clear metric gain, since the re-projected color triplane already carries reference appearance information; its benefit is qualitative (global tone transfer).
Limitations
- Ground regions (road, sidewalk) are not recovered well in geometry and appearance.
- Transient masks are learned unsupervised and can wrongly mask hard-to-model but static areas, causing color inconsistency.
- Still requires at least double the training time of vanilla 3DGS to reach comparable rendering quality.
Experimental setup
Datasets: Phototourism (Brandenburg Gate, Sacre Coeur, Trevi Fountain)
Benchmarks: Phototourism Brandenburg Gate Phototourism Sacre Coeur Phototourism Trevi Fountain
Metrics: PSNR SSIM LPIPS training time (hours) inference speed (FPS)
Baselines: 3DGS NeRF-W Ha-NeRF CR-NeRF GS-W SWAG
Builds on: 3D Gaussian Splatting NeRF-W CR-NeRF EG3D FSGS
Closest papers by meaning
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.92)
- Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.90)
- Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment (CVPR 2025 · similarity 0.89)
- GS-IR: 3D Gaussian Splatting for Inverse Rendering (CVPR 2024 · similarity 0.89)
- PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- WildCAT3D: Appearance-Aware Multi-View Diffusion in the Wild (NeurIPS 2025 · similarity 0.89)
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.