Radiance field-based 3D reconstruction and potential applications

Radiance field-based 3D reconstruction has emerged as a transformative research direction due to its remarkable efficiency and quality. This paper presents a systematic analysis of representation models, reconstruction methodologies, and future applications in this field. We start from an overview of multi-view 3D reconstruction tasks, then focus on the key issue: how to represent 3D content effectively. Radiance fields are highlighted for their flexibility and representational completeness. Distinguished from the existing review literature, we adopt a multi-dimensional comparison between neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS) to develop a unified and in-depth understanding of the radiance field-based approach. Beyond the initial goal of novel view synthesis (NVS), recent breakthroughs in geometry extraction are summarized. Finally, we explore potential applications across areas such as robot localization and mapping, virtual reality, physical simulation, and stereo display. Empowered by the flexible 3D representation within the radiance field-based paradigm, the latest advancements strive to push the boundaries and overcome long-standing bottlenecks in related domains.