Self-supervised Dense Depth Prediction in Monocular Endoscope Video for 3D Liver Surface Reconstruction

Abstract

In this paper, we propose a self-supervised monocular depth prediction method which combines traditional multi-view stereo method and fully convolutional network to predict the depth map in monocular endoscopic video and achieve 3D dense reconstruction of liver surface. We adopt the sparse data generated by COLMAP supervision signal as the training data, and integrate the attention model in the fully convolutional network to effectively extract the channel features to improve the accuracy of depth prediction. Taking into account the problem of insufficient supervision ability of sparse data, the projection transformation of two images within a certain range is carried out to make up for the missing supervision points. Experimental results show that this method has achieved good results in the depth prediction of monocular endoscopic video and has good applicability to the whole liver.