Automated keypoint definition for surgical instruments

Abstract

Pose estimation in stereo vision relies on robust correspondences between image features. Instead of expensive detection and matching, regression networks can be trained to localize distinctive features directly, but manual feature annotation can bias results. This study introduces methods to increase performance by automatically selecting keypoints from 2D feature detectors. Across five surgical instruments, we compare keypoints derived from SIFT (Scale-invariant feature transform) and ORB (Oriented FAST and Rotated BRIEF). Subsequently, the number of selected keypoints with regard to the performance of subsequent keypoint regression is assessed. To avoid errors from manual annotation and to enable efficient scalability, all experiments use purely synthetic, automatically generated datasets.