Core Concepts
Endo-4DGS introduces a real-time endoscopic dynamic reconstruction approach using 3D Gaussian Splatting for accurate and efficient scene reconstruction.
Abstract
Introduction:
Importance of accurate 3D reconstruction in endoscopic surgery.
Challenges posed by complex endoscopic scenes.
Advancements in Endoscopic 3D Reconstruction:
Utilization of Deep Neural Networks (DNNs) and Neural Radiance Fields (NeRFs).
Previous studies leveraging NeRF for tissue deformation modeling.
Challenges Addressed by Endo-4DGS:
Slow rendering speeds and suboptimal localization accuracy faced by NeRF-based methods.
Introduction of 3D Gaussian Splatting as an alternative for fast inference and superior representation.
Methodology:
Incorporation of temporal dynamics with lightweight MLPs and Gaussian deformation fields.
Utilization of Depth-Anything for robust depth estimation and Gaussian initialization.
Results:
Validation on surgical datasets showcasing real-time rendering, computational efficiency, and high accuracy compared to existing methods.
Conclusion:
Proposal of Endo-4DGS as a solution for real-time, high-fidelity reconstruction of deformable tissues in endoscopic surgery.
Stats
Neural Radiance Fields (NeRF)-based methods have slow inference speed, prolonged training, and inconsistent depth estimation.
Our approach has been validated on two surgical datasets, demonstrating real-time rendering, computational efficiency, and remarkable accuracy.