Non-Volumetric Dynamic Reconstruction With Adaptive Correction

Single-view RGBD 3D reconstruction, particularly for dynamic scenes, remains a challenging task in computer vision. Traditional non-rigid reconstruction methods often struggle with large topological changes, leading to errors in model area addition and even tracking failure. Additionally, these methods rely on volumetric structures, resulting in significant memory and time overhead. In this paper, we propose a non-volumetric dynamic reconstruction approach that introduces an adaptive canonical model correction. Our method addresses non-rigid alignment failures caused by large topological changes and maintains only one-dimensional surface elements (surfels) for fusion processing. This reduces the need for high-performance servers and enhances the information content of reconstructed elements. Experimental results demonstrate that our approach efficiently handles complex and rapid topological changes. Our method provides a robust and efficient solution for real-time dynamic reconstruction of non-rigid scenes.