Consistent 6D pose estimation via geometry-guided object pose graph construction

The process of determining the position and orientation of an object in 3D space with respect to the camera is known as 6D pose estimation. This is a fundamental problem in computer vision that has numerous real-world applications. Most of the existing approaches in the literature focus solely on the object itself, without considering its relationships with other objects within the scene. Therefore, in this thesis, we propose a novel method for refining the monocular 6D object pose at the instance level of multiple objects by leveraging geometric information about the relationships of neighboring objects. Our experimental results demonstrate that our architecture, which is trained exclusively on synthetic images, can produce satisfactory results within the synthetic domain. However, when tested on real-world images, it does not deliver the same level of performance.