MrNeRF (@janusch_patas)
2025-11-20 | โค๏ธ 86 | ๐ 12
Gaussian See, Gaussian Do: Semantic 3D Motion Transfer from Multiview Video
Contributions: โข We introduce a new problem setup for semantic 3D motion transfer, where motion from a multiview video of a dynamic source object is adapted to a static 3D Gaussian Splatting (3DGS) target, even across vastly different categories, without requiring predefined skeletal correspondences or structural similarity.
โข We propose a novel view-aware motion embedding strategy, optimized collaboratively through an anchor-based interpolator. This approach balances global coherence with view-dependent details, ensuring motion fidelity while significantly accelerating convergence.
โข We develop a robust 4D reconstruction pipeline that refines noisy, artifact-prone supervision videos into high-quality, temporally stable dynamic 3D reconstructions.
โข We introduce a benchmark for semantic 3D motion transfer, evaluating diverse cross-category motion transfers between structurally different 3D objects.
๐ ์๋ณธ ๋งํฌ
๋ฏธ๋์ด
