Research map / 3D vision / Gaussian splatting
Free360: Layered Gaussian Splatting for Unbounded 360-Degree View Synthesis from Extremely Sparse and Unposed Views
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.
Free360 reconstructs unbounded 360-degree scenes from 3-9 unposed sparse views by combining a layered 2D Gaussian representation with layer-specific bootstrap optimization and an iterative reconstruction-generation loop using ViewCrafter with uncertainty-aware training.
Problem
Novel view synthesis and 3D reconstruction in unbounded 360-degree scenes from extremely sparse (3-9) unposed views, where stereo point clouds are noisy and video diffusion outputs are multi-view inconsistent.
Approach
Partition the scene into front and back Gaussian layers initialized from DUSt3R point clouds, refine each layer with tailored bootstrap losses (photometric, normal, depth, inpainting for back layer), and iteratively fuse reconstruction with ViewCrafter generation while weighting supervision by epistemic uncertainty from perturbed-condition ensembles.
Key results
- On Mip-NeRF 360 with 3 views, Free360 reaches 16.20 PSNR / 0.378 SSIM / 0.499 LPIPS, outperforming ViewCrafter (15.63 PSNR) and InstantSplat (14.10 PSNR).
- On Tanks and Temples with 4 views, Free360 reaches 14.69 PSNR vs. 13.56 for ViewCrafter and 13.03 for InstantSplat.
- Surface reconstruction F1 on four Tanks and Temples scenes is 0.423 at 4 views vs. 0.284 for 2DGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Free360 instead of ViewCrafter | Mip-NeRF 360, 3-view novel view synthesis · Mip-NeRF 360, 3 views | PSNR: 15.63 -> 16.20 | Helped |
| Use Free360 instead of ViewCrafter | Tanks and Temples, 4-view novel view synthesis · Tanks and Temples, 4 views | PSNR: 13.56 -> 14.69 | Helped |
| Add multi-layer representation (ablation) | Free360 baseline · Mip-NeRF 360, 3 views | PSNR: 16.16 -> 16.56 | Helped |
| Add bootstrap optimization (ablation) | Free360 with multi-layer representation · Mip-NeRF 360, 3 views | PSNR: 16.56 -> 16.65 | Helped |
| Add iterative fusion (ablation) | Free360 with bootstrap optimization · Mip-NeRF 360, 3 views | PSNR: 16.65 -> 16.81 | Helped |
| Add uncertainty-aware training (ablation) | Free360 with iterative fusion · Mip-NeRF 360, 3 views | PSNR: 16.81 -> 16.95 | Helped |
| Use Free360 instead of 2DGS | Tanks and Temples surface reconstruction · Tanks and Temples, 4 scenes | F1-score: 0.284 -> 0.423 | Helped |
Limitations
- Fails on scenes with extensive repetitive texture where stereo reconstruction breaks down
- Depends on the quality of DUSt3R poses and point clouds; Flower scene excluded because DUSt3R fails with sparse views
- Inherits multi-view inconsistency from the video diffusion prior (ViewCrafter)
Experimental setup
Datasets: Mip-NeRF 360 Tanks and Temples
Benchmarks: Mip-NeRF 360 Tanks and Temples
Metrics: PSNR SSIM LPIPS F1-score
Baselines: FSGS InstantSplat ZeroNVS ViewCrafter 2DGS
Builds on: 2D Gaussian Splatting DUSt3R ViewCrafter InstantSplat RAIN-GS
Closest papers by meaning
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.90)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.90)
- FatesGS: Fast and Accurate Sparse-View Surface Reconstruction Using Gaussian Splatting with Depth-Feature Consistency (AAAI 2025 · similarity 0.90)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.89)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.89)
- Binocular-Guided 3D Gaussian Splatting with View Consistency for Sparse View Synthesis (NeurIPS 2024 · 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.