Research map / 3D vision / Gaussian splatting

One World, Dual Timeline: Decoupled Spatio-Temporal Gaussian Scene Graph for 4D Cooperative Driving Reconstruction

NeurIPS 2026 · 3D vision › Gaussian splatting · approach: Monocular dynamic scene reconstruction · 0 citations (OpenAlex)

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 DUST, a decoupled spatio-temporal Gaussian scene graph that shares canonical agent Gaussians but uses separate vehicle and infrastructure pose timelines, plus static-anchor pose correction and pose regularization, for 4D cooperative driving reconstruction on V2X-Seq.

Problem

Reconstructing dynamic agents from vehicle-infrastructure cooperative driving data is hampered by temporal asynchrony between independently clocked cameras, which single-pose-per-agent Gaussian scene graphs cannot represent, causing ghosting.

Approach

Share a canonical Gaussian set per agent across sources but assign each source its own pose trajectory sampled at its true capture time, with static-vehicle anchor-based label correction and a pose-regularized joint optimization; a theorem claims single-timeline formulations incur an irreducible loss quadratic in velocity and time offset.

Key results

• Dynamic-area scene reconstruction PSNR 26.11 vs 22.91 for PVG (about 3.2 dB gain) on V2X-Seq full average.
• Scene reconstruction dynamic-area FVD drops from 63.7 (OMNIRE) to 39.7 (37.7% reduction).
• On High Crowd subset, FVD 47.4 vs OMNIRE's 89.3; ablation shows removing dual timelines drops dynamic PSNR from 25.12 to 23.35.

Surprising findings

• Removing the drift loss causes only a slight PSNR drop, though it is argued to be vital for dynamic stability.

Limitations

• Does not explicitly handle shape changes of deformable agents such as pedestrians under asynchronous observations.
• Theory is local and linearized, assuming constant velocity and locally rigid motion over the time offset.

Experimental setup

Datasets: V2X-Seq

Benchmarks: V2X-Seq

Metrics: PSNR SSIM LPIPS FVD RAFT-EPE

Baselines: 3DGS DeformableGS StreetGS OMNIRE CRUISE PVG

Builds on: OMNIRE Gaussian Scene Graph 3D Gaussian Splatting Slerp Hungarian method

Closest papers by meaning

• FRUC: Feedforward Dynamic Scene Reconstruction from Uncalibrated Collaborative Driving Views (NeurIPS 2026 · similarity 0.90)
• DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes (CVPR 2024 · similarity 0.88)
• DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input (AAAI 2025 · similarity 0.88)
• DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes (CVPR 2025 · similarity 0.88)
• Flux4D: Flow-based Unsupervised 4D Reconstruction (NeurIPS 2025 · similarity 0.87)
• DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving (NeurIPS 2025 · similarity 0.86)