Real-time large-scale reconstruction from multi-camera rig video streams

Achieve real-time reconstruction of complete, large-scale 3D scenes directly from video streams captured by multi-camera rigs.

Background

The paper surveys on-the-fly 3D Gaussian Splatting approaches and notes that most existing methods operate on monocular inputs, which leads to incomplete coverage and missing cross-view consistency. While multi-camera rigs inherently provide overlapping views and improved scene coverage, prior work has not delivered a real-time solution that reconstructs complete, large-scale scenes from rig video streams.

The authors explicitly state that achieving real-time reconstruction from multi-camera rig streams remains open and challenging, motivating their proposed framework designed for large-scale outdoor scenes without explicit calibration.