everyone
since 04 Oct 2024">EveryoneRevisionsBibTeXCC BY 4.0
In the field of scene reconstruction with moving objects, recent studies have utilized 3D Gaussian Splatting (3DGS) for spatial representation. This method typically relies on camera poses and point clouds obtained through the Structure-from-Motion (SfM) algorithm. However, in scenes captured with monocular viewpoints and containing moving objects in each frame, the SfM algorithm struggles to obtain accurate camera poses and points clouds. As a result, it often either removes point clouds of dynamic objects or fails to find camera poses for each frame, thereby leading to sub-optimal rendering of dynamic scenes. We propose a novel approach, Initial-Free Monocular Dynamic Gaussian Splatting (IF-MoDGS) which does not require precomputed camera poses and point clouds in dynamic scenes with moving objects. Our approach estimates camera poses using the static background, separated from dynamic objects by a motion mask, and generates point clouds specifically for the dynamic objects. To handle dynamic objects, we define a canonical space and apply deformation to link it with each viewpoint and timestamp. Then, to improve quality in complex spatio-temporal scenes, we utilize a high-dimensional feature loss and an annealing frequency loss. Extensive experimental results demonstrate that our method can effectively render dynamic scenes without relying on precomputed camera poses and point clouds, achieving the state-of-the-art performance in dynamic scene rendering tasks using a monocular camera. Our project will be available at:https://anonymous.4open.science/w/IF-MODGS-67F5/