MrNeRF (@janusch_patas)
2025-10-31 | โค๏ธ 126 | ๐ 20
Motion4D: Learning 3D-Consistent Motion and Semantics for 4D Scene Understanding
โข We propose Motion4D, a model that integrates 2D priors from foundation models into a dynamic 3D Gaussian Splatting representation. This achieves consistent motion and semantic modeling from monocular videos.
โข We design a two-part iterative optimization framework comprising:
- Sequential optimization: updates motion and semantic fields in consecutive stages to maintain local consistency.
- Global optimization: jointly refines all attributes to ensure long-term coherence.
โข We introduce iterative motion refinement using 3D confidence maps and adaptive resampling to enhance dynamic scene reconstruction. Semantic refinement corrects 2D semantic inconsistencies through iterative updates with SAM2.
โข Our Motion4D significantly outperforms both 2D foundation models and existing 3D methods in tasks such as video object segmentation, point-based tracking, and novel view synthesis.
๋ฏธ๋์ด
