ColonAdapter: Geometry Estimation Through Foundation Model Adaptation for Colonoscopy
ColonAdapter adapts large 3D geometric foundation models for colonoscopy geometry estimation using self-supervised fine-tuning from monocular videos. It addresses clinical challenges like textureless regions by introducing a Detail Restoration Module and specialized loss functions. The method successfully improves camera pose, depth, and point map accuracy on synthetic and real datasets without needing ground truth intrinsic parameters.