Depth-Regularized 3D Gaussian Splatting for Robust Endoscopic Reconstruction in Feature-Scarce Environments.

3D reconstruction in endoscopic imaging can significantly improve diagnostic accuracy, enhance surgical planning, and provide realistic training environments for clinicians. However, existing reconstruction methods like Neural Radiance Fields and standard 3D Gaussian Splatting (3DGS) struggle in gastrointestinal (GI) environments, where narrow viewpoints and uniform textures hinder reliable 3D modeling. To overcome these challenges, we propose an advanced 3DGS-based framework that integrates a deep learning-based Structure-from-Motion (SfM) technique with sophisticated depth regularization. Our method leverages Super-point and Superglue within the SfM process to extract and match features robustly from GI tract scenes, facilitating accurate camera pose estimation and effective initialization of 3D Gaussians. We further refine the reconstruction by aligning monocular depth predictions from a pre-trained Depth-Anything-V2 model with SfM-derived depth using a scale-offset adjustment, enforced by an L loss. In addition, hard depth regularization, imposed via a Huber loss, ensures precise placement of 3D Gaussians, while globallocal depth normalization preserves both fine local details and overall structural consistency. Extensive experiments on multiple endoscopic datasets demonstrate that our approach delivers enhanced reconstruction quality and reduced artifacts.