Efficiently Reconstructing Dynamic Scenes One D4RT at a Time

October 14, 2026

-a unified transformer jointly infers depth, spatiotemporal correspondence, and full camera parameters from a single video

-enables 4D reconstruction

-state-of-the-art 3D reconstruction methods such as VGGT cannot establish correspondence for dynamic portions of a scene

-a global self-attention encoder first transforms the input video into a global scene-representation latent

-the decoder can be independently queried for the 3D position of any given 2D point from

t_{\\text{source}}

t_{\\text{target}}

in camera coordinate

t_{\\text{cam}}

-3D models: better spatial understanding, trained on image datasets

-does spatial understanding over time lead to good 4D/video understanding?

-4D models: better temporal understanding, trained on video datasets

-open question: can 4D models become as strong at spatial understanding as 3D models?