A review of dynamic scene reconstruction based on neural representation

Abstract

Dynamic scene reconstruction holds significant research value in the fields of computer vision and virtual reality. Recent advancements in neural representation technologies have facilitated rapid progress in this task. Over the past four years, methods based on neural radiance fields and 3D Gaussian splatting have been proposed, achieving remarkable results. However, the large number of literature presents a challenge for individuals to comprehensively track comprehensive relevant works. To address this issue, typical work for dynamic scene reconstruction based on neural representation was summarized, categorizing them into methods based on neural radiance fields and 3D Gaussian splatting. Furthermore, representative datasets were highlighted and common evaluation metrics for algorithms were summarized. Finally, the persistent challenges in current methodologies were discussed and potential directions for future development trends were proposed.

Keywords

• dynamic reconstruction
• neural representation
• neural radiance field
• 3D Gaussian splatting