Research map / 3D vision / Gaussian splatting
HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene
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 HAIF-GS, a sparse anchor-driven dynamic 3D Gaussian Splatting framework with anchor filtering, self-supervised induced-flow deformation, and hierarchical anchor densification for monocular dynamic scene reconstruction.
Problem
Reconstructing dynamic 3D scenes from monocular video with 3DGS, where existing methods suffer from redundant Gaussian updates, insufficient motion supervision, and weak modeling of fine-grained non-rigid deformation.
Approach
Drives Gaussian deformation through sparse motion anchors filtered by a learned dynamic confidence score; induces anchor flow self-supervisedly via multi-frame feature aggregation with a cycle-consistency loss (no flow labels); and hierarchically densifies anchors in high-variance-motion regions with parent-to-child transformation propagation.
Key results
- On NeRF-DS, HAIF-GS reports mean PSNR 24.63 and MS-SSIM 0.9014 versus SC-GS 24.05 and 0.8848 and 4DGS 24.18.
- On D-NeRF, HAIF-GS reports mean PSNR 42.00 and SSIM 0.997 versus SC-GS 41.65 and 0.993.
- Ablations show removing the induced flow module, cycle loss, anchor filter, or hierarchical densification each degrades performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HAIF-GS instead of NeRF-DS | Dynamic scene reconstruction · NeRF-DS mean | PSNR: 23.60 -> 24.63 | Helped |
| Use HAIF-GS instead of SC-GS | Dynamic scene reconstruction · NeRF-DS mean | MS-SSIM: 0.8848 -> 0.9014 | Helped |
| Use HAIF-GS instead of SC-GS | Dynamic scene reconstruction · D-NeRF mean | LPIPS: 0.009 -> 0.010 | Hurt |
| Remove Induced Flow-Guided Deformation (ablation) | HAIF-GS · Ablation study | PSNR | Hurt |
| Remove cycle consistency loss (ablation) | HAIF-GS · Ablation study | PSNR | Hurt |
| Remove Anchor Filter (ablation) | HAIF-GS · Ablation study | PSNR | Hurt |
| Remove Hierarchical Anchor Densification (ablation) | HAIF-GS · Ablation study | PSNR | Hurt |
Limitations
- Memory footprint remains non-trivial for highly complex motion due to anchor proliferation.
- Implicit self-supervised flow may be sensitive to significant occlusions.
Experimental setup
Datasets: NeRF-DS D-NeRF
Benchmarks: NeRF-DS D-NeRF
Metrics: PSNR SSIM MS-SSIM LPIPS
Baselines: 3DGS TiNeuVox HyperNeRF NeRF-DS 4DGS Deformable 3DGS SC-GS SP-GS
Builds on: 3D Gaussian Splatting SC-GS Deformable 3DGS linear blend skinning
Closest papers by meaning
- MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.92)
- H3D-DGS: Exploring Heterogeneous 3D Motion Representation for Deformable 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes (CVPR 2024 · similarity 0.91)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.91)
- Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 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.