Flux4D

Overview

Flux4D is a flow-based unsupervised 4D reconstruction framework for large-scale dynamic scenes. It directly predicts 3D Gaussians and their motion dynamics from raw LiDAR and camera data using only photometric losses and an 'as static as possible' regularization. It reconstructs scenes within seconds, scales to large datasets, and generalizes to unseen environments. It outperforms existing unsupervised methods and achieves competitive performance with supervised approaches on outdoor driving datasets.

Key features

• ▸Unsupervised 4D reconstruction from raw LiDAR and camera data
• ▸Directly predicts 3D Gaussians and their motion dynamics
• ▸Reconstructs dynamic scenes within seconds
• ▸Scales effectively to large datasets
• ▸Generalizes well to unseen environments
• ▸Handles complex dynamic scenarios in the wild
• ▸Processes full-resolution images (≥ 1920 x 1080)
• ▸Enables controllable camera simulation without labels