Abstract:
The field of 3D reconstruction has undergone significant advancements with the
integration of machine learning techniques, enabling more efficient and accurate modeling
of complex environments. This research focuses on leveraging cutting-edge
methodologies, such as 3D Gaussian Splatting, to address challenges in real-time rendering
and dynamic scene reconstruction. Traditional methods for dynamic scene rendering often
fall short in maintaining high-quality and real-time performance, especially for complex,
moving environments. In contrast, Gaussian Splatting employs probabilistic primitives to
represent 3D point clouds, offering a balance between computational efficiency and visual
quality. This study extends Progressive Gaussian Splatting to dynamic environments by
introducing a framework that ensures temporal coherence and real-time performance. The
proposed methodology employs a hybrid geometric representation, progressive
propagation for Gaussian refinement, and deformation fields encoded via multi-resolution
voxel grids to capture motion. Evaluations on synthetic and real-world datasets
demonstrate significant improvements in rendering quality and temporal coherence,
achieving state-of-the-art results in metrics such as PSNR (41.99), SSIM (0.995), and
LPIPS (0.011) for synthetic dataset and increased PSNR by 1.79 and SSIM by 0.046 for
real world’s hypernerf dataset. The research provides new insights into the practical
application of Gaussian Splatting in dynamic environments, opening avenues for enhanced
virtual reality (VR), augmented reality (AR), and robotics applications.
