DC-Forward: Robust FeedForward Reconstruction from In-the-wild Driving-Scene Videos

Overview

Driving-scene videos captured in the wild contain foreground occlusions and reflections that corrupt feedforward reconstruction. DC-Forward is a 4D Gaussian Splatting (4DGS) reconstruction model built to stay robust under these conditions.

Contributions

In-the-wild training data at scale

Built a 4DGS-based perturbation pipeline that uses text-guided segmentation and semantic opacity refinement to extract foreground occlusions from real videos, with configurable camera parameters. With it, generated 4,000+ in-the-wild driving-scene videos from nuScenes, Waymo, and other public datasets.

Patch-level reliability

Proposed a 4DGS reconstruction architecture that predicts patch-level reliability for geometry and appearance. Each patch is routed to geometry estimation, appearance modeling, or both, and patches that support neither are discarded.

Results

Resolves foreground-occlusion and reflection artifacts in in-the-wild driving scenes, improving PSNR by approximately 2 dB on average over state-of-the-art methods (DGGT, STORM, and Instant-NuRec).

Context

Project leader and first author of the paper, which is ready for submission to CVPR 2027. Advisor: Prof. Zhiwen Fan, PHAI Lab, Texas A&M University. Aug 2026 – present.