Abstract
Modern 3D engines and graphics pipelines require mesh as a memory-efficientrepresentation, which allows efficient rendering, geometry processing, textureediting, and many other downstream operations. However, it is still highlydifficult to obtain high-quality mesh in terms of structure and detail frommonocular visual observations. The problem becomes even more challenging fordynamic scenes and objects. To this end, we introduce Dynamic Gaussians Mesh(DG-Mesh), a framework to reconstruct a high-fidelity and time-consistent meshgiven a single monocular video. Our work leverages the recent advancement in 3DGaussian Splatting to construct the mesh sequence with temporal consistencyfrom a video. Building on top of this representation, DG-Mesh recovershigh-quality meshes from the Gaussian points and can track the mesh verticesover time, which enables applications such as texture editing on dynamicobjects. We introduce the Gaussian-Mesh Anchoring, which encourages evenlydistributed Gaussians, resulting better mesh reconstruction through mesh-guideddensification and pruning on the deformed Gaussians. By applyingcycle-consistent deformation between the canonical and the deformed space, wecan project the anchored Gaussian back to the canonical space and optimizeGaussians across all time frames. During the evaluation on different datasets,DG-Mesh provides significantly better mesh reconstruction and rendering thanbaselines.