Uni4R Marries Optimal Transport and ODEs for Unified 4D Reconstruction

zhenjun_zhao · x · 2026-08-13

The paper proposes Uni4R, a framework that unifies 4D reconstruction and point tracking by learning continuous velocity fields through the synergy of Optimal Transport (OT) and Ordinary Differential Equations (ODE).

Existing methods typically rely on heuristic interpolations, lacking kinematic coherence. The continuous velocity field proposed in Uni4R acts as a kinematic prior that mutually benefits both tasks. At its core, the Flow Matching Guided Decoder (FMGD) extracts anchor features capturing global dynamics and decodes velocity information at arbitrary timestamps.

Related event: Uni4R Unifies 4D Reconstruction and Tracking via OT and ODE(2 posts)→

Original post →

More from Research

Research channel →