Research map / 3D vision / Gaussian splatting
HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 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 HiCoM, an online 3DGS framework for streamable multi-view dynamic scenes that combines perturbation-smoothed initial Gaussians, hierarchical region-shared motion, and continual refinement with Gaussian merging, plus parallel multi-frame training.
Problem
Online reconstruction of dynamic scenes from multi-view video streams is limited by training time, rendering speed, and storage/transmission cost, and 3DGS overfits with sparse views and grows Gaussians unchecked.
Approach
Model frame-to-frame motion explicitly with hierarchical regions of Gaussians that share translation and rotation parameters across multiple spatial scales (only non-empty regions get parameters), pair this with noise-perturbed initial 3DGS training, and keep newly cloned Gaussians by removing an equal number of low-opacity ones.
Key results
- On N3DV, HiCoM reaches 31.17 dB PSNR with 0.7 MB/frame storage and 6.7 s training per frame, versus 30.15 dB, 7.6 MB, and 8.2 s for 3DGStream.
- On Meet Room, HiCoM achieves 26.73 dB with 0.4 MB per frame and 3.9 s training, and the 4-frame parallel variant trains in 1.0 s per frame at 26.25 dB.
- Storage after the initial frame is under 10% of 3DGStream's per-frame storage, with roughly 17-20% training time reduction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · N3DV | PSNR: 30.15 -> 31.17 | Helped |
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · Meet Room | PSNR: 25.96 -> 26.73 | Helped |
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · N3DV | Storage (MB/frame): 7.6/7.8 -> 0.7/0.9 | Helped |
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · Meet Room | Storage (MB/frame): 4.0/4.1 -> 0.4/0.6 | Helped |
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · N3DV | Training time (s/frame): 8.2 -> 6.7 | Helped |
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · Meet Room | Training time (s/frame): 4.7 -> 3.9 | Helped |
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · N3DV | Rendering speed (FPS): 210 -> 274 | Helped |
| Use HiCoM rather than 3DGStream | Online dynamic-scene reconstruction · Meet Room | Rendering speed (FPS): 212 -> 284 | Helped |
| Add perturbation smoothing to initial 3DGS learning (ablation) | HiCoM · Coffee Martini, Flame Steak, and Discussion | PSNR: 26.61 -> 28.04; 32.52 -> 32.87; 26.19 -> 26.69 | Helped |
| Add hierarchical coherent motion learning (ablation) | HiCoM · Coffee Martini, Flame Steak, and Discussion | PSNR: 25.71 -> 28.04; 27.33 -> 32.87; 21.61 -> 26.69 | Helped |
| Add continual refinement with new Gaussians (ablation) | HiCoM · Coffee Martini, Flame Steak, and Discussion | PSNR: 27.87 -> 28.04; 31.73 -> 32.87; 26.51 -> 26.69 | Helped |
| Increase motion-learning steps from 50 to 100 (ablation) | HiCoM · N3DV | PSNR: 30.78 -> 31.17 | Helped |
Surprising findings
- Parallel training with 2 or 4 frames sometimes slightly improves PSNR over sequential training on some scenes.
- Using the standard 3DGS initial frame (no perturbation) gives only PSNR comparable to 3DGStream, so the initial representation quality matters beyond the motion module.
Limitations
- Experiments are limited to indoor scenes; generalization to outdoor scenes is unverified.
- Reconstruction quality may decline over time due to error accumulation in online learning.
- The initial 3DGS representation remains a crucial bottleneck that is not fully addressed.
Experimental setup
Datasets: N3DV Meet Room PanopticSports Particle-NeRF
Benchmarks: N3DV Meet Room PanopticSports Particle-NeRF
Metrics: PSNR training time per frame rendering FPS storage per frame
Baselines: StreamRF 3DGStream Naive 3DGS Dynamic 3DGS 4D-GS Particle-NeRF
Builds on: 3D Gaussian Splatting 3DGStream 4D-GS StreamRF
Closest papers by meaning
- H3D-DGS: Exploring Heterogeneous 3D Motion Representation for Deformable 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- Motion Matters: Compact Gaussian Streaming for Free-Viewpoint Video Reconstruction (NeurIPS 2025 · similarity 0.92)
- Instant Gaussian Stream: Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting (CVPR 2025 · similarity 0.92)
- HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene (NeurIPS 2025 · similarity 0.92)
- SfM-Free 3D Gaussian Splatting via Hierarchical Training (CVPR 2025 · similarity 0.91)
- MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
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.