Research map / 3D vision / Gaussian splatting
KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences
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 KeyGS, a keyframe-centric 3DGS framework that gets rough poses from sequential COLMAP SfM on subsampled keyframes plus SLERP, then jointly refines poses and Gaussians with coarse-to-fine frequency-aware densification, without depth or matching models.
Problem
Novel view synthesis and camera pose estimation from monocular image sequences without accurate poses; existing joint-refinement NeRF/3DGS methods are slow (hours to days) and error-prone under high-frequency signals.
Approach
Use fast sequential SfM on uniformly subsampled keyframes with SLERP interpolation for initial poses, then jointly optimize SE(3) pose offsets and 3DGS while applying a coarse-to-fine Gaussian-blur schedule to densification (frequency-aware densification) to avoid over-splitting and pose local minima.
Key results
- On Tanks and Temples, KeyGS averages 33.01 PSNR / 0.94 SSIM / 0.04 LPIPS, about 2 dB above CF3DGS (31.28).
- On CO3Dv2 five sequences, KeyGS averages 30.12 PSNR vs 26.32 for CF3DGS and 25.14 for Nope-NeRF, with training in about 10 minutes versus 2 hours (CF3DGS) and 30 hours (Nope-NeRF).
- Ablation: removing coarse-to-fine frequency-aware densification drops PSNR from 33.01 to 19.70 (no components) or 18.13 (only absolute gradient and anisotropy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KeyGS instead of CF3DGS | Tanks and Temples · Tanks and Temples mean | PSNR: 31.28 -> 33.01 | Helped |
| Use KeyGS instead of CF3DGS | Tanks and Temples · Tanks and Temples mean | LPIPS: 0.09 -> 0.04 | Helped |
| Use KeyGS instead of CF3DGS | Tanks and Temples · Tanks and Temples mean | RPE_t: 0.041 -> 0.020 | Helped |
| Use KeyGS instead of CF3DGS | CO3DV2 · CO3DV2 average | PSNR: 26.32 -> 30.12 | Helped |
| Use KeyGS instead of CF3DGS | CO3DV2 · CO3DV2 average | LPIPS: 0.31 -> 0.15 | Helped |
| Use KeyGS instead of CF3DGS | CO3DV2 · CO3DV2 average | RPE_t: 0.217 -> 0.031 | Helped |
| Add coarse-to-fine frequency-aware densification (ablation) | KeyGS · Tanks and Temples | PSNR: 19.70 -> 33.01 | Helped |
| Add absolute gradient (ablation) | KeyGS with C2F and anisotropy regularization · Tanks and Temples | PSNR: 31.17 -> 33.01 | Helped |
| Add anisotropy regularization (ablation) | KeyGS with C2F and absolute gradient · Tanks and Temples | PSNR: 32.94 -> 33.01 | Helped |
| Use rough poses with joint refinement instead of 3DGS with COLMAP poses (ablation) | Tanks and Temples · Tanks and Temples | PSNR: 30.77 -> 33.01 | Helped |
Surprising findings
- Naive joint refinement with 3DGS without the coarse-to-fine component performs far worse (PSNR 19.70) than 3DGS given COLMAP poses (30.77), showing that the densification schedule is essential.
- Using exhaustive-mode COLMAP poses with 3DGS underperforms KeyGS's joint refinement from rough sequential poses (30.77 vs 33.01 PSNR).
Limitations
- Keyframe subsampling interval of 10 is not robust for outdoor scenes, so a interval of 5 is used, which reduces the speedup.
- Downsampling images for faster SfM can fail in some scenes due to lack of robust feature points.
Experimental setup
Datasets: Tanks and Temples CO3Dv2
Benchmarks: Tanks and Temples CO3Dv2
Metrics: PSNR SSIM LPIPS RPE_t RPE_r ATE
Baselines: BARF CF3DGS Nope-NeRF SC-NeRF 3DGS with COLMAP poses
Builds on: 3D Gaussian Splatting BARF Joint TensoRF COLMAP AbsGS
Closest papers by meaning
- SfM-Free 3D Gaussian Splatting via Hierarchical Training (CVPR 2025 · similarity 0.93)
- 4D Gaussian Splatting in the Wild with Uncertainty-Aware Regularization (NeurIPS 2024 · similarity 0.91)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.91)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.91)
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos (NeurIPS 2025 · similarity 0.90)
- No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 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.