Research map / 3D vision / Gaussian splatting
ReCon-GS: Continuum-Preserved Guassian Streaming for Fast and Compact Reconstruction of Dynamic Scenes
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 ReCon-GS, a storage-aware streaming 3DGS framework that uses multi-level anchor Gaussians with grid-based farthest point sampling, periodic re-hierarchization with intra-level deformation inheritance, and tunable anchor density for online dynamic scene reconstruction.
Problem
Online free-viewpoint video reconstruction of dynamic scenes suffers from slow per-frame optimization, error accumulation in streaming deformation, and high per-frame storage, with no way to trade fidelity against memory at deployment.
Approach
Decompose per-frame Gaussian motion into three hierarchical levels of anchor-driven rigid transformations placed by grid-based farthest point sampling; periodically re-hierarchize anchors and inherit deformation parameters only from matched anchors within the same level, while tuning anchor density to trade storage against quality.
Key results
- On N3DV, ReCon-GS reaches 32.66 dB PSNR with 0.44 MB/frame storage (0.40 MB without initial frame) and 6.4 s training, outperforming HiCoM (32.08 dB, 0.69 MB) and QUEEN-l (32.19 dB).
- On Technicolor, ReCon-GS achieves 33.83 dB PSNR at 0.82 MB storage and 207 FPS, versus E-D3DGS 33.24 dB at 1.54 MB.
- Storage can be adjusted by anchor density: from 0.14 MB/frame (1/96 density, 32.11 dB) to 0.63 MB/frame (1/16 density, 32.72 dB) on N3DV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the 3-level hierarchy rather than 2 levels (ablation) | ReCon-GS · N3DV | PSNR (dB): 32.422 -> 32.662 | Helped |
| use the 3-level hierarchy rather than 2 levels (ablation) | ReCon-GS · Meet Room | PSNR (dB): 30.20 -> 30.84 | Helped |
| increase anchor density from 1/96 to 1/16 (ablation) | ReCon-GS · N3DV | PSNR (dB): 32.11 -> 32.72 | Helped |
| increase Phase 1 deformation training from 50 to 150 steps (ablation) | ReCon-GS · N3DV | PSNR (dB): 32.40 -> 32.76 | Helped |
| increase Phase 1 deformation training from 50 to 150 steps (ablation) | ReCon-GS · N3DV | Train (sec): 5.20 -> 7.66 | Hurt |
| set Phase 2 densification to 100 steps rather than 50 or 150 (ablation) | ReCon-GS · N3DV | Storage (MB): 0.46; 0.44; 0.49 | Helped |
| train Frame 0 for 10,000 rather than 2,500 steps (ablation) | ReCon-GS · N3DV | PSNR (dB): 31.03 -> 32.63 | Helped |
| increase Frame 0 training from 2,500 to 10,000 steps (ablation) | ReCon-GS · N3DV | Slope: 0.0077 -> 0.0016 | Helped |
| use 3-level rather than 4-level hierarchy (ablation) | ReCon-GS · N3DV | PSNR (dB): 32.658 -> 32.662 | No clear effect |
Surprising findings
- Table 3 shows the initial frame storage for the 2,500-iteration setting (8.96 MB) being larger than for 10,000 iterations (11.12 MB), while incremental storage drops sharply with longer frame-0 training.
Limitations
- Not specifically optimized for monocular datasets or specialized reconstruction scenarios.
- Struggles with dynamic allocation of Gaussians for real-time object emergence and dissolution.
- Offline methods still use less storage than ReCon-GS due to implicit motion representation.
Experimental setup
Datasets: N3DV Meet Room PanopticSports Technicolor
Benchmarks: N3DV Meet Room PanopticSports Technicolor
Metrics: PSNR SSIM LPIPS storage (MB) training time (s) rendering FPS
Baselines: 4DGS STG SaRO-GS Swift4D SplineGS Dynamic 3DGS StreamRF 3DGStream 4DGC QUEEN-l HiCoM IGS-l
Builds on: 3D Gaussian Splatting Scaffold-GS HiCoM 3DGStream 4DGS
Closest papers by meaning
- HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene (NeurIPS 2025 · similarity 0.89)
- MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.88)
- Instant Gaussian Stream: Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Motion Matters: Compact Gaussian Streaming for Free-Viewpoint Video Reconstruction (NeurIPS 2025 · similarity 0.88)
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.