Decoupled Dynamic Spatio-Temporal Neural Radiance Fields for Enhanced 3D Scene Synthesis

Novel view synthesis of dynamic scenes represents a critical research direction in computer vision and computer graphics. In particular, when addressing complex dynamic scenes, effectively separating the static background from the dynamic foreground and generating high-quality synthetic views remains a significant challenge. This paper introduces a Dynamic SpaceTime Decoupling Neural Radiance Field (DDST-NeRF) method. By modeling static and dynamic scenes separately and integrating spatio-temporal modeling with dynamic weighting fusion, DDSTNeRF enables the generation of novel views and perspectives at any time from monocular videos. Experimental results demonstrate that DDST-NeRF performs exceptionally well on multiple datasets for dynamic scenes, especially in handling rigid motion scenarios, where it achieves significantly better PSNR and SSIM indices compared to existing methods. However, there is still room for improvement in non-rigid motion scenarios. The research presented in this paper offers a novel solution for dynamic scene view synthesis and provides a promising direction for future studies.