Self-Supervised Micro-Baseline Binocular Endoscope Depth Estimation Method With Domain Adaptation

Computing depth information from two-dimensional (2D) endoscopic images can provide three-dimensional (3D) depth perception for the surgeon and the surgical robot. Compared with the monocular depth estimation algorithm, the binocular depth estimation paradigm has superiority in supplying stable 3D information. For the binocular endoscope system, reducing the baseline distance can decrease the diameter of the binocular endoscope. However, achieving accurate binocular depth estimation with the micro-baseline binocular endoscope presents challenges. Therefore, we propose a micro-baseline binocular endoscope depth estimation method (MB-BEDE) based on domain adaptation. Furthermore, we employ the view-synthesis strategy to construct our self-supervised depth estimation network to eliminate the need for hard-to-obtain depth ground truth during training. Results on two self-collected micro-baseline binocular endoscope datasets show that MB-BEDE can achieve accurate depth estimation with 2.60mm and 3.23mm errors on the respective datasets, demonstrating the superiority of our method in micro-baseline binocular endoscope depth estimation.