NeRF dynamic scene reconstruction based on motion, semantic information and inpainting

Abstract

In this work, we address the inherent limitations of Neural Radiance Field (NeRF) in synthesizing novel viewpoints within dynamic environments, particularly those compromised by moving objects. Such scenarios frequently yield reconstructions of suboptimal quality, characterized by blurriness and the presence of artifacts, which significantly undermines the fidelity of synthetic scenes. This limitation significantly restricts the potential applications of NeRF in autonomous driving contexts, such as scene editing, high-precision map construction, and related functionalities. To overcome these challenges, we propose a novel NeRF-based approach tailored to address the complexities associated with moving objects in monocular driving scenarios. The proposed approach combines optical flow analysis and semantic information to precisely detect and localize moving objects. This was then followed by an inpainting technique that guides the NeRF reconstruction process, effectively mitigating the adverse impacts of dynamic elements within the scene. Our model is further enhanced by incorporating depth and semantic data to refine the training process. We validate the efficacy of our approach through comprehensive experimentation on both synthetic and real-world driving datasets, as well as on challenging self-recorded realistic driving scenes. Our method achieves a performance improvement of up to 13% compared to previous state-of-the-art methods. Additionally, we verify the efficacy of our approach through comprehensive ablation analyses. Both the quantitative and qualitative results demonstrate the superiority especially in dynamic driving scenes, advancing the potential applications in autonomous driving contexts.

Our code and self-collected data are available at https://github.com/GandalfTGrey/Nerf-KBS.git.