Deep Learning-Based 6-DoF Object Pose Estimation With Synthetic Data: A Case Study in Underwater Environments

In this thesis we aim to address the image based 6-DoF pose estimation problem, or 3D pose estimation problem, for Autonomous Underwater Vehicles (AUVs). The results of the object pose estimation will be used, for example, to estimate the global location of the AUV or to approach more accurately the underwater infrastructures. Actually, an autonomous robot or a team of autonomous robots need accurate location skills to safely and effectively move within an underwater environment, where communications are sparse and unreliable, and to accomplish high-level tasks such as: underwater exploration, mapping of the surrounding environment, multi-robot conveyance and many other multi-robot problems. Several state-of-the-art approaches will be analysed and tested on real datasets. Collecting underwater images and providing them with an accurate ground-truth estimate of the object's pose is an expansive and extremely time-consuming activity To this end, we addressed the problem using only synthetic datasets. In fact, it was not possible to use the standard datasets used in the analyzed papers, since they are datasets with objects and conditions very different from those in which the AUVs operate. Hence, we exploited an unpaired image-to-image translation network is employed to bridge the gap between the rendered and the real images, producing photorealistic synthetic training images. Promising preliminary results confirm the goodness of the made choices.