Research map / 3D vision / Gaussian splatting
COLMAP-Free 3D 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 CF-3DGS, which jointly estimates camera poses and reconstructs a 3D Gaussian Splatting scene by sequentially growing Gaussians over video frames, without COLMAP preprocessing.
Problem
Novel view synthesis from images without known or SfM-precomputed camera poses, which is hard for NeRFs under large camera motion and slow to train.
Approach
Uses explicit 3D Gaussians: a local 3DGS estimates the relative pose between adjacent frames by optimizing a learnable SE(3) transform of frozen Gaussians, and a global 3DGS progressively grows with each incoming frame, with densification tied to frame arrival.
Key results
- On Tanks and Temples, CF-3DGS reaches mean PSNR 31.28 / SSIM 0.93 / LPIPS 0.09 versus 26.34 PSNR for Nope-NeRF.
- On CO3D-V2 with large 360-degree motion, CF-3DGS reduces mean ATE substantially compared with Nope-NeRF (e.g., 0.008 vs 0.053 in one scene) and trains in about 2 hours versus about 30 hours for Nope-NeRF.
- Removing progressive growing drops mean PSNR from 31.28 to 22.57, showing the sequential growing is essential.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CF-3DGS instead of Nope-NeRF | Novel view synthesis on Tanks and Temples · Tanks and Temples | PSNR: 26.34 -> 31.28 | Helped |
| use CF-3DGS instead of Nope-NeRF | Novel view synthesis on Tanks and Temples · Tanks and Temples | LPIPS: 0.39 -> 0.09 | Helped |
| use CF-3DGS instead of Nope-NeRF | Camera pose estimation on Tanks and Temples · Tanks and Temples | RPEt: 0.080 -> 0.041 | Helped |
| use CF-3DGS instead of Nope-NeRF | Camera pose estimation on Tanks and Temples · Tanks and Temples | RPEr: 0.038 -> 0.069 | Hurt |
| use CF-3DGS instead of Nope-NeRF | Novel view synthesis on CO3D V2 scene 110 · CO3D V2, scene 110 | PSNR: 26.86 -> 29.69 | Helped |
| use CF-3DGS instead of Nope-NeRF | Camera pose estimation on CO3D V2 scene 110 · CO3D V2, scene 110 | RPEr: 1.966 -> 0.401 | Helped |
| add progressive growing (ablation) | CF-3DGS on Tanks and Temples · Tanks and Temples | PSNR: 22.57 -> 31.28 | Helped |
| add progressive growing (ablation) | CF-3DGS on Tanks and Temples · Tanks and Temples | RPEt: 0.116 -> 0.041 | Helped |
| add depth loss to the photometric objective (ablation) | CF-3DGS on Tanks and Temples · Tanks and Temples | PSNR: 28.94 -> 31.28 | Hurt |
| add depth loss to the photometric objective (ablation) | CF-3DGS on Tanks and Temples · Tanks and Temples | RPEt: 0.039 -> 0.041 | No clear effect |
Surprising findings
- Adding depth loss did not help and slightly reduced novel view synthesis quality (mean PSNR 28.94 vs 31.28).
- Pose accuracy on Tanks and Temples is only comparable to Nope-NeRF rather than clearly better, with worse relative rotation error.
Limitations
- Requires ordered inputs such as video streams; does not handle unordered image collections.
- Relative pose estimation relies solely on photometric loss, giving poorer rotation error than Nope-NeRF on some scenes.
Experimental setup
Datasets: Tanks and Temples CO3D-V2
Benchmarks: Tanks and Temples CO3D-V2
Metrics: PSNR SSIM LPIPS RPE translation RPE rotation ATE training time
Baselines: Nope-NeRF BARF NeRFmm SC-NeRF 3DGS with COLMAP poses
Builds on: 3D Gaussian Splatting NeRF Nope-NeRF BARF
Closest papers by meaning
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.93)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.93)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.93)
- SfM-Free 3D Gaussian Splatting via Hierarchical Training (CVPR 2025 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.92)
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.