Real-Time Coarse-to-Fine Depth Estimation on Stereo Endoscopic Images With Self-Supervised Learning

Fast and accurate depth estimation is an essential task in computer-assisted surgery and robotics, especially for endoscopic and microscopic procedures. We propose a real-time stereo matching model using a staged, coarse-to-fine architecture to estimate disparity from medical stereo camera data with self-supervised learning. Our model processes images with a resolution of $1280 \times 1024$ pixels beyond 60 fps, with similar accuracy to the semi-global matching algorithm, and does not require any ground truth depth for training. We evaluated our model on two stereo endoscopic datasets from the literature. A mean absolute error below 1.5 mm and root mean square error below 1.9 mm were identified.