Dynamic Neural Radiance Fields Based on Camera Pose Estimation Using Monocular Vision

Gengpei Fu, Fan Li, Mingyun He, Ping Kuang2025 · DOI

Abstract

As an emerging view synthesis technique, Neural Radiance Fields (N eRF) leverages the advantages of implicit neural representations to enable high-fidelity novel view synthesis of complex 3D scenes, driving the development of technologies such as autonomous driving and virtual reality, and has emerged as a research hotspot in recent years. However, traditional NeRF encounters bottlenecks in dynamic scene modeling due to its static scene assumption and reliance on precise camera poses, struggling to meet the demands of complex dynamic scenarios. Addressing the issue of unstable camera pose estimation in dynamic scenes, this paper designs a monocular-vision-based dynamic NeRF algorithm. By adopting a j oint optimization framework for static-dynamic radiance fields, it simultaneously optimizes scene representations and camera parameters, eliminating the reliance on high-precision camera poses. Experiments demonstrate that under complex motion scenarios where precise poses are not pre-acquired, this method improves view synthesis accuracy by 3.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> compared to existing approaches dependent on precise poses, providing an effective solution for high-fidelity view synthesis in dynamic scenes.
BibTeX
@misc{fu2025dynamic,
  author = {Gengpei Fu and Fan Li and Mingyun He and Ping Kuang},
  title = {{Dynamic Neural Radiance Fields Based on Camera Pose Estimation Using Monocular Vision}},
  year = {2025},
  doi = {10.1109/icmlca66850.2025.11336467},
}

Loading related content...

Figure