WS-SfMLearner: self-supervised monocular depth and ego-motion estimation on surgical videos with unknown camera...

Area of Science:

• Computer Vision
• Medical Imaging
• Robotics

Background:

• Accurate depth estimation is crucial for image-guided surgery.
• Creating ground truth depth maps for surgical videos is challenging due to illumination and sensor noise.
• Current self-supervised methods require known camera intrinsic parameters, often unavailable in surgical settings.

Purpose of the Study:

• To develop a self-supervised system for joint depth, ego-motion, and intrinsic parameter estimation in surgical videos.
• To overcome the limitation of unknown or missing camera intrinsic parameters in surgical environments.
• To provide a comprehensive solution for depth estimation in complex surgical video scenarios.

Main Methods:

• Developed a self-supervised depth and ego-motion estimation framework.
• Incorporated a cost volume-based auxiliary supervision module for intrinsic parameter prediction.
• Evaluated the system on a public dataset for simultaneous prediction of depth, camera pose, and intrinsic parameters.

Main Results:

• Significantly improved accuracy in ego-motion and depth prediction compared to methods using known intrinsics.
• Enhanced accuracy of camera parameter estimation, including intrinsic parameters, through cost volume-based supervision.
• Demonstrated robust performance even with unknown camera intrinsic parameters.

Conclusions:

• The proposed self-supervised system effectively estimates depth, ego-motion, and intrinsic parameters without prior knowledge of intrinsics.
• Outperformed baseline techniques in complex surgical video scenarios.
• Offers a robust solution with broader implications for advancing image-guided surgery systems.