Scale-preserving shape reconstruction from monocular endoscope image sequences by supervised depth learning.

Area of Science:

• Computer Vision
• Medical Imaging
• 3D Reconstruction

Background:

• 3D shape reconstruction from images is gaining traction.
• Existing methods often produce relative depth maps with uncertain scales.
• Accurate scale-preserving 3D reconstruction is crucial for various applications.

Purpose of the Study:

• To propose a novel method for reconstructing scale-preserving 3D shapes from monocular endoscope image sequences.
• To develop an absolute depth prediction network for accurate scale estimation.
• To validate the method on both simulated and real endoscope data.

Main Methods:

• A dataset of synchronized RGB images and depth maps was generated using an endoscope simulator.
• A supervised depth prediction network was trained to estimate depth maps from RGB images, minimizing loss against ground-truth depth.
• The predicted depth map sequences were aligned to reconstruct the final 3D shape.

Main Results:

• The trained network successfully estimated absolute depth maps from monocular endoscope images.
• The alignment of predicted depth maps enabled the reconstruction of scale-preserving 3D shapes.
• The proposed method demonstrated effectiveness when applied to a real endoscope image sequence.

Conclusions:

• The developed absolute depth prediction network effectively addresses the scale ambiguity in monocular endoscope-based 3D reconstruction.
• This method offers a promising approach for creating accurate 3D models from endoscopic imagery.
• The technique has potential applications in medical diagnostics and surgical planning.