CNN Approach for Monocular Depth Estimation: Ear Case Study

The realization of medical devices to assist the surgeon in autologous auricular reconstruction (personalized surgical guides) requires the use of depth maps images obtained from 3D scans of the patient's profile. In order to make the process of ear geometry acquisition and depth map realization faster, more comfortable for the patient and easily accessible by hospital staff, this work proposes a system able to create depth maps from a single RGB image. The proposed approach involves the integration of tools based on convolutional neural networks to build a modular system capable of isolating the ear from the profile, creating the corresponding depth map, and refining it to correct inaccuracies. The system was trained and tested using a database of human profile images and corresponding depth maps made available by the University of Notre Dame. To evaluate the result in this preliminary study, standard metrics such as mean square error and structural similarity were used, yielding results suitable for the targeted application.