Deep neural networks (DNNs) are increasingly employed across diverse fields of applied science, particularly in areas like computer vision and image processing, where they enhance the performance of instruments. Various advanced coherent imaging techniques, including digital holography, leverage different deep architectures like convolutional neural networks (CNN) or Vision Transformers (ViT). These architectures enable the extraction of diverse metrics such as autofocusing reconstruction distance or 3D position determination, facilitating applications in automated microscopy and phase image restitution. In this work, we propose a hybrid approach utilizing an adapted version of the GedankenNet model, coupled with a UNet-like model, for the purpose of accessing micro-objects 3D pose measurements. These networks are trained on simulated holographic datasets. Our approach achieves an accuracy of 98% in inferring the 3D poses. We show that a GedankenNet can be used as a regression tool and is faster than a Tiny-ViT (TViT) model. Overall, integrating deep neural networks into digital holographic microscopy and 3D computer micro-vision holds the promise of significantly enhancing the robustness and processing speed of holograms for precise 3D position inference and control, particularly in micro-robotics applications.