MrNeRF (@janusch_patas)
2025-12-09 | โค๏ธ 75 | ๐ 15
Neural Mesh Gaussians for Robust Monocular Reconstruction of Dynamic Objects
Contributions: โข Our method reconstructs dynamic objects from a single monocular video under challenging camera setups, applicable to category-agnostic objects.
โข We leverage LRM to prepare per-frame coarse geometry prior and propose a deformation MLP to build frame correspondence, faithfully matching the dynamic information of the input video.
โข We propose a novel mesh Gaussian structure that offers higher fidelity appearance, lower memory burden, and increased training speed. Additionally, we introduce two constraints for mesh Gaussians to avoid view-overfitting.
๋ฏธ๋์ด
