Research map / 3D vision / Gaussian splatting

Flux4D: Flow-based Unsupervised 4D Reconstruction

NeurIPS 2025 · 3D vision › Gaussian splatting · approach: Monocular dynamic scene reconstruction · 3 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.

Flux4D predicts 3D Gaussians with per-Gaussian velocities from LiDAR-initialized multi-sensor driving data, trained unsupervised with photometric/depth losses and an "as static as possible" velocity penalty across many scenes, enabling fast generalizable 4D reconstruction.

Problem

Annotation-free, scalable and fast 4D reconstruction of large dynamic driving scenes, separating static background from moving actors without per-scene optimization or pre-trained perception priors.

Approach

A feed-forward sparse 3D U-Net predicts Gaussian residuals and velocities from LiDAR-initialized Gaussians, trained only with photometric, SSIM, depth losses plus an L2 velocity penalty across many scenes; the static-dynamic decomposition emerges from data-driven priors, augmented by iterative gradient-feedback refinement and velocity-based loss reweighting.

Key results

• On PandaSet 1s interpolation, Flux4D achieves 21.99 dynamic-only PSNR and 23.84 full-image PSNR, versus 18.62 for Flux4D-base and 17.65 for STORM, in 1.8s reconstruction time.
• On WOD NVS (DrivingRecon setup), Flux4D reaches 26.62 PSNR and 0.82 SSIM full-image, +5.99 dB PSNR over DrivingRecon (20.63).
• Flux4D's 8-second full-log dynamic PSNR is 21.94, above unsupervised EmerNeRF (18.65) and DeSiRe-GS (19.76), and close to supervised NeuRAD (22.99).

Measured effects reported in the paper

Surprising findings

• Higher-order (degree 2 and 3) polynomial motion gave no noticeable gain over the constant-acceleration model in the authors' experiments.
• Removing the velocity regularization drastically raises velocity RMSE (0.157 to 0.532) even though dynamic-only depth MAE slightly improves (1.44 vs 1.63).

Limitations

• Flow estimation for highly dynamic actors with complex motion is challenging.
• Iterative approach for long-horizon reconstruction creates visible inconsistencies at transition points.
• Assumes a pinhole camera model and clean LiDAR data, limiting rolling-shutter or noisy-sensor use.

Experimental setup

Datasets: PandaSet Waymo Open Dataset Argoverse 2

Benchmarks: PandaSet Waymo Open Dataset

Metrics: PSNR SSIM LPIPS depth MAE normalized velocity RMSE EPE3D Acc5 Acc10 angular error EPE-3way

Baselines: EmerNeRF DeSiRe-GS DepthSplat L4GM STORM DrivingRecon NeuRAD StreetGS G3R NSFP FastNSF

Builds on: 3D Gaussian Splatting G3R EmerNeRF STORM

Closest papers by meaning

• 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.90)
• Motion4D: Learning 3D-Consistent Motion and Semantics for 4D Scene Understanding (NeurIPS 2025 · similarity 0.89)
• Instant4D: 4D Gaussian Splatting in Minutes (NeurIPS 2025 · similarity 0.88)
• Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos (CVPR 2025 · similarity 0.88)
• Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.88)
• IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer (NeurIPS 2026 · similarity 0.88)