CAD-Based RGB-Only 6D Pose Estimation: Enhancing Precision for Unseen and Textureless Objects

6D object pose estimation is a recurrent and crucial task, particularly for robotic applications such as pick-and-place. While the research community has developed solutions with varying levels of precision, reaching peak performance with textured objects and networks fine-tuned on predefined sets, a significant performance gap remains in more challenging settings. This thesis addresses this specific gap by focusing on a highly challenging yet industrially relevant setting: zero-shot 6D pose estimation of unseen, textureless objects using solely RGB visual information from a single view (without depth cues) and the CAD model of the object. Starting from an existing pipeline, this work investigates its limitations and proposes targeted enhancements to increase both precision and robustness. The contributions of this thesis are threefold: first, we analyze all the possible stages of a pose estimation pipeline, assessing the impact of removing or modifying them; second, we improve the existing pipeline by both evaluating various feature-matching approaches to bridge the real-to-synthetic domain gap (selecting the most effective strategy to improve pose estimation) and introducing an alternative method to handle pose ambiguities; third, we present custom-acquired datasets designed to rigorously validate the pipeline's applicability in realistic scenarios and to mitigate errors along the depth axis.