With the development of deep learning, many edge devices adopt monocular depth estimation (MDE) to produce reliable 3D RGB-D data due to its low power and low costs compared to an RGB-D sensor [1]. In a user domain, the MDE's deep neural network has to be re-trained for the domain adaptation [2], [3]. However, the conventional training system requires expensive labeled dataset collection, which is obtained by a high-cost RGB-D sensor with additional raw data preprocessing [3] such as depth alignment, depth synchronization, and depth inpainting. As a result, the proposed processor performs the unsupervised learning-based MDE system, which does not rely on additional sensors and preprocessing for data collection for the training.