Research paper on reconstructing dynamic 3D scenes from multi-camera video that wasn't synchronized. The paper proposes Phase-DGS, which swaps per-frame timestamps for a "semantic phase" describing where each frame sits in the motion cycle.
The problem: Dynamic Gaussian Splatting assumes all cameras are perfectly synced. The abstract says that's rarely true in practice, and the usual fixes need special hardware or environment-specific assumptions.
The approach: Frames showing the same motion state get the same phase, whichever camera recorded them and whenever, which the authors say establishes correspondence across unsynced cameras.
Results: Tested under temporal offsets, frame drops and camera freezes, with up to +8.5 dB PSNR over baselines, per the abstract. It also plugs into existing backbones like RealTime-4DGS and FreeTimeGS.
Status: Listed as accepted to NeurIPS 2026, with a project page linked in the comments.
Sign in to comment.
TL;DR of the paper, generated automatically.
Research paper on reconstructing dynamic 3D scenes from multi-camera video that wasn't synchronized. The paper proposes Phase-DGS, which swaps per-frame timestamps for a "semantic phase" describing where each frame sits in the motion cycle.