Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild

4 hours ago

Lift4D is a test-time optimization framework for reconstructing complete dynamic objects from monocular video, integrating visual cues with data-driven priors.
It improves on prior methods by using causal latent conditioning for temporally consistent per-frame 3D predictions, initializing a deformable 3D Gaussian Splatting representation.
The framework sculpts the representation via occlusion-aware optimization, recovering visible surface details and completing unobserved regions with a view-conditioned diffusion prior.
It demonstrates superior performance on challenging in-the-wild sequences with severe occlusions and non-rigid motion compared to prior 4D reconstruction methods.

Hasty Briefsbeta