Research map / 3D vision / Gaussian splatting
SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video
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.
SplineGS proposes a COLMAP-free dynamic 3D Gaussian Splatting framework that models each dynamic Gaussian's trajectory with cubic Hermite splines whose control points are pruned adaptively, enabling real-time rendering from monocular videos.
Problem
Novel view and time synthesis of dynamic scenes from in-the-wild monocular videos, where COLMAP camera estimates are unreliable and existing dynamic 3DGS deformation models are slow or inflexible.
Approach
Represent each dynamic 3D Gaussian's mean trajectory with a cubic Hermite spline over learnable control points, prune control points per Gaussian by a least-squares error threshold (MACP), and jointly optimize camera poses and intrinsics with photometric and geometric consistency instead of SfM preprocessing.
Key results
- On the NVIDIA dataset (SfM-free setting), SplineGS reaches 27.21 average PSNR and 0.053 LPIPS at 400 FPS, versus 25.38 PSNR for RoDynRF (SfM-free) and 26.61 for MoSca (SfM-free).
- In the novel view and time synthesis experiment, SplineGS achieves 25.92 PSNR, 0.098 LPIPS and 0.703 tOF, outperforming RoDynRF (21.58 PSNR) and STGS (15.72 PSNR).
- Ablation: replacing MAS with a Bezier curve gives 27.19 PSNR and an MLP gives 23.51 PSNR; removing the photometric consistency loss drops PSNR to 17.49.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SplineGS rather than MoSca for SfM-free reconstruction | NVIDIA dataset · NVIDIA dataset (SfM-free) | PSNR: 26.61 -> 27.21 | Helped |
| Use SplineGS rather than MoSca for SfM-free reconstruction | NVIDIA dataset · NVIDIA dataset (SfM-free) | LPIPS: 0.069 -> 0.053 | Helped |
| Use SplineGS rather than RoDynRF for SfM-free reconstruction | NVIDIA dataset · NVIDIA dataset (SfM-free) | FPS: 0.45 -> 400 | Helped |
| Use SplineGS for synthesis at unseen intermediate timestamps | NVIDIA dataset · NVIDIA dataset, novel view and time synthesis | PSNR: 23.36 -> 25.92 | Helped |
| Use SplineGS for synthesis at unseen intermediate timestamps | NVIDIA dataset · NVIDIA dataset, novel view and time synthesis | tOF: 0.921 -> 0.703 | Helped |
| Use cubic Hermite MAS rather than MLP deformation (ablation) | SplineGS · NVIDIA dataset | PSNR: 23.51 -> 27.21 | Helped |
| Use cubic Hermite MAS rather than grid-based deformation (ablation) | SplineGS · NVIDIA dataset | PSNR: 25.48 -> 27.21 | Helped |
| Use cubic Hermite MAS rather than a Bézier curve (ablation) | SplineGS · NVIDIA dataset | PSNR: 27.19 -> 27.21 | Helped |
| Adaptively prune control points with MACP instead of fixing Nc = 4 (ablation) | SplineGS · NVIDIA dataset | PSNR: 26.62 -> 27.21 | Helped |
| Adaptively prune control points with MACP instead of fixing Nc = Nf (ablation) | SplineGS · NVIDIA dataset | PSNR: 27.08 -> 27.21 | Helped |
| Add photometric consistency loss Lpc (ablation) | SplineGS · NVIDIA dataset | PSNR: 17.49 -> 27.21 | Helped |
| Add geometric consistency loss Lgc (ablation) | SplineGS · NVIDIA dataset | PSNR: 26.33 -> 27.21 | Helped |
Surprising findings
- Removing the photometric consistency loss collapses PSNR to 17.49, showing camera estimation is essential without SfM.
- Ex4DGS and STGS, designed for multi-view settings, struggle with geometry alignment over time when trained on monocular video.
Limitations
- Relies on pre-computed motion masks (Track Anything) and metric depth (UniDepth) and 2D tracks (CoTracker) as priors.
- Evaluated mainly on the NVIDIA dataset; DAVIS evaluation is qualitative only.
- Fixed-degree polynomials and MLPs are shown inferior, but the MACP threshold epsilon is a hand-set hyperparameter.
Experimental setup
Datasets: NVIDIA dynamic dataset DAVIS
Benchmarks: NVIDIA dynamic dataset DAVIS
Metrics: PSNR LPIPS tOF FPS training time
Baselines: DynNeRF MonoNeRF STGS SC-GS D3DGS 4DGS SP-GS Casual-FVS Kinematic Fields E-D3DGS CTNeRF D-NPC
Builds on: 3D Gaussian Splatting STGS D3DGS CoTracker UniDepth
Closest papers by meaning
- MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors (ICLR 2025 · similarity 0.91)
- Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling (AAAI 2025 · similarity 0.90)
- MotionGS: Exploring Explicit Motion Guidance for Deformable 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- 4D Gaussian Splatting in the Wild with Uncertainty-Aware Regularization (NeurIPS 2024 · similarity 0.90)
- KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences (AAAI 2025 · similarity 0.90)
- World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026 · 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.