A feedforward 4D Gaussian Splatting reconstruction model for in-the-wild driving-scene videos that predicts patch-level reliability for geometry and appearance, removing foreground-occlusion and reflection artifacts.
Publications
Papers and preprints.
*Equal contribution. †Corresponding author.
Conference
DC-Forward: Robust FeedForward Reconstruction from In-the-wild Driving-Scene Videos
WorldCorner: Evaluating Driving World Models for Safety-Critical Event Generation
The first benchmark dedicated to evaluating whether driving world models can generate safety-critical events, shifting the evaluation target from perceptual realism to event-level fidelity.
CrashRecast: Agentic Crash Reconstruction from Accident Reports via World Models
An agentic workflow that infers crash trajectories from real accident reports, and a pipeline that turns those trajectories into photorealistic 4D Gaussian Splatting scenes.
Journal
Reconstructing 3D Voxels from Ultra-sparse-view 2D X-ray Images Using a Dual-domain Balanced Network
A U-Net framework that reconstructs CBCT voxels from only 2–4 X-ray views, reducing radiation dose by approximately 50% while achieving SSIM above 0.94.