Research map / 3D vision / Gaussian splatting
Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling
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 EDGS, which models monocular dynamic scenes with a sparse anchor-grid, an unsupervised time-mask MLP that selects deformable anchors, and RBF-kernel-based sparse motion for Gaussian offsets, yielding faster rendering with higher quality.
Problem
Rendering dynamic scenes from monocular videos with deformable Gaussian splatting is slow because it produces many redundant Gaussians and queries time-variant attributes for all of them, including static regions that also jitter.
Approach
Combine a Scaffold-GS-style sparse anchor grid with an unsupervised time-mask MLP that filters static anchors so only deformable anchors are decoded by time-variant MLPs, and compute dense Gaussian offset motion from anchor motion via an RBF kernel on anchor and offset features.
Key results
- On HyperNeRF vrig, EDGS reaches 25.7 dB PSNR, 0.860 MS-SSIM and 117 FPS, versus 25.0 dB and 13 FPS for Deformable 3DGS and 25.2 dB and 34 FPS for 4DGS.
- On NeRF-DS, the method reports mean PSNR 24.39 and SSIM 0.8873 across scenes, with the highest mean PSNR and SSIM among compared methods.
- Ablation on three NeRF-DS scenes: removing the time-mask drops speed from about 150 FPS to about 125 FPS and PSNR on Sieve from 27.12 to 26.18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add anchor-grid initialization (ablation) | EDGS · NeRF-DS As | FPS: 27 -> 152 | Helped |
| add anchor-grid initialization (ablation) | EDGS · NeRF-DS Bell | FPS: 23 -> 147 | Helped |
| add anchor-grid initialization (ablation) | EDGS · NeRF-DS Sieve | FPS: 17 -> 151 | Helped |
| add time-mask MLP (ablation) | EDGS · NeRF-DS As | PSNR: 26.02 -> 26.65 | Helped |
| add time-mask MLP (ablation) | EDGS · NeRF-DS Bell | PSNR: 25.78 -> 26.01 | Helped |
| add time-mask MLP (ablation) | EDGS · NeRF-DS Sieve | PSNR: 26.18 -> 27.12 | Helped |
| replace rigid offset transformation with RBF kernel (ablation) | EDGS · NeRF-DS As, Bell, and Sieve | PSNR: 23.51 -> 26.65; 24.21 -> 26.01; 23.95 -> 27.12 | Helped |
| replace KNN offset derivation with RBF kernel (ablation) | EDGS · NeRF-DS As, Bell, and Sieve | PSNR: 25.75 -> 26.65; 25.53 -> 26.01; 25.68 -> 27.12 | Helped |
| replace cosine-similarity offset derivation with RBF kernel (ablation) | EDGS · NeRF-DS As, Bell, and Sieve | PSNR: 26.31 -> 26.65; 25.74 -> 26.01; 26.57 -> 27.12 | Helped |
| replace cosine-similarity offset derivation with RBF kernel (ablation) | EDGS · NeRF-DS As, Bell, and Sieve | SSIM: 0.8928 -> 0.9015; 0.8807 -> 0.9203; 0.9155 -> 0.9271 | Mixed |
Surprising findings
- Removing the anchor-grid initialization drops rendering speed to around 17-27 FPS, showing that Gaussian count is the main speed bottleneck.
- Cosine similarity offset derivation performs similarly to the RBF kernel but is VRAM-consuming.
Limitations
- Anchor positions are frozen after initialization from COLMAP points, so anchors are not updated during training.
- KNN-based offset derivation degrades when moving objects separate after collision, motivating the kernel choice but not fully solved.
Experimental setup
Datasets: NeRF-DS HyperNeRF
Benchmarks: NeRF-DS HyperNeRF vrig
Metrics: PSNR SSIM LPIPS MS-SSIM FPS training time
Baselines: 3D-GS TiNeuVox HyperNeRF NeRF-DS 4DGS SCGS Deformable 3DGS Nerfies FFDNeRF
Builds on: 3D Gaussian Splatting Scaffold-GS Deformable 3DGS SCGS
Closest papers by meaning
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- 1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering (NeurIPS 2025 · similarity 0.92)
- Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis (CVPR 2024 · similarity 0.92)
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes (CVPR 2024 · similarity 0.92)
- MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors (ICLR 2025 · similarity 0.92)
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene 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.