4D Gaussian Reconstruction from Monocular Dynamic Videos
2025.02 -- 2025.08
Research Intern
A4x, Hangzhou, China
- Foundation-model priors: Use VGGT to initialize geometry and camera poses; combine SAM 2 video segmentation, CoTracker point tracking, and Depth Anything depth estimation to constrain monocular dynamic Gaussian reconstruction and recover geometry, appearance, and motion.
- Motion fields and dynamic rendering: Fit Gaussians' global SE(3) motion and local non-rigid deformation in stages using local residual motion control points. Adaptively add/prune these motion control points to refine local motion and deformation, enabling dynamic novel-view rendering and dense 3D trajectory extraction.