Research map / 3D vision / Gaussian splatting
ProDyG: Progressive Dynamic Scene Reconstruction via Gaussian Splatting from Monocular Videos
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 ProDyG, an online SLAM-based system that tracks monocular or RGB-D video with flow-based motion masks refined by SAM2 and reconstructs static and dynamic parts progressively using Gaussian splatting with Motion Scaffolds.
Problem
Online dynamic 3D scene reconstruction from RGB or RGB-D streams that requires global pose and map consistency, detailed appearance, and scalability, which existing dynamic SLAM, offline, and feed-forward methods do not jointly satisfy.
Approach
Disentangles static and dynamic content inside a Splat-SLAM backend: residual-flow motion masks (refined with SAM2 point prompts) suppress dynamic pixels during bundle adjustment, and dynamic Gaussians are progressively attached to Motion Scaffolds nodes that are updated as new keyframes arrive, with an added motion-mask loss that keeps dynamic opacity out of static regions.
Key results
- On the iPhone dataset NVS, ProDyG (RGB-D, online) reaches PSNR 17.87 and SSIM 0.643, exceeding offline Shape of Motion (17.43 / 0.591) but below MoSca (18.44 / 0.666).
- Tracking on Bonn RGB-D Dynamic averages ATE 3.29 cm, competitive with WildGS-SLAM (2.94 cm), and clearly better than Splat-SLAM (10.6 cm).
- RGB-only ProDyG achieves PSNR 15.41 on iPhone, the first online RGB-only dynamic reconstruction method in the comparison, though with lower quality than RGB-D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProDyG's motion-agnostic camera tracking | RGB-D SLAM on Bonn RGB-D Dynamic Dataset · Bonn RGB-D Dynamic Dataset (Avg.) | ATE RMSE [cm]: 2.94 -> 3.29 | Hurt |
| Use ProDyG's motion-agnostic camera tracking | RGB-D SLAM on TUM RGB-D Dataset · TUM RGB-D Dataset (Avg.) | ATE RMSE [cm]: 1.63 -> 1.89 | Hurt |
| Run online reconstruction and tracking with RGB-D input instead of offline MoSca reconstruction | ProDyG · iPhone Dataset | PSNR: 18.44 -> 17.87 | Hurt |
| Run online reconstruction and tracking with RGB-D input instead of offline MoSca reconstruction | ProDyG · iPhone Dataset | SSIM: 0.666 -> 0.643 | Hurt |
| Run online reconstruction with offline tracking and RGB-D input instead of Shape of Motion | ProDyG · iPhone Dataset | PSNR: 17.43 -> 17.65 | Helped |
| Run online reconstruction with offline tracking and RGB-D input instead of Shape of Motion | ProDyG · iPhone Dataset | SSIM: 0.591 -> 0.634 | Helped |
| Run online reconstruction and tracking with RGB-D input instead of online reconstruction with precomputed camera poses | ProDyG · iPhone Dataset | PSNR: 17.65 -> 17.87 | Mixed |
| Run online reconstruction and tracking with RGB-only input instead of RGB-D input | ProDyG · iPhone Dataset | PSNR: 17.87 -> 15.40 | Hurt |
Surprising findings
- WildGS-SLAM, a static-focused method with soft uncertainty masks, outperforms ProDyG on tracking, which the authors attribute to aggressive suppression of moving shadows and DINOv2 features.
Limitations
- Struggles with objects that move mostly outside the viewing frustum and reappear, leading to regularization-dominated deformations.
- Requires per-scene test-time optimization, so it is not real-time.
- Cannot handle large novel viewpoint changes without generative priors.
Experimental setup
Datasets: Bonn RGB-D Dynamic Dataset TUM RGB-D Dataset iPhone Dataset
Benchmarks: Bonn RGB-D Dynamic Dataset TUM RGB-D Dataset iPhone Dataset
Metrics: ATE RMSE PSNR SSIM LPIPS
Baselines: Splat-SLAM WildGS-SLAM MegaSaM DynaMoN DROID-SLAM Shape of Motion DynOMo MoSca Gaussian Marbles ORB-SLAM2 DDN-SLAM
Builds on: Splat-SLAM Motion Scaffolds DROID-SLAM SAM2 CoTracker3 RaDe-GS
Closest papers by meaning
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.91)
- World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.90)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.90)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.90)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.89)
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.89)
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.