6D object pose estimation is a computer vision task popularly used in, among others, virtual reality or robotics applications. Currently, it is still an open problem with a lot of space for improvement, especially in the category-level field.

Moreover, a category-level 6D pose estimation algorithms could prove useful in applications developed by Aivero AS, which would use them for vision and robotics contexts in real-life scenarios.

This master thesis focuses on exploring the problem by researching the latest developed solutions, public datasets, as well as detection and evaluation methods. Based on the carried out research, FS-Net had been identified as one of the most accurate open-source networks, and therefore was used throughout the project. First, this work attempted to reproduce the results by training the network on NOCS-Real dataset. Next, it was attempted to improve the models produced by network with more focus put on 'mug' category.

Although this work did not succeed in improving the results, the n\degree m cm and IoU metrics have been reproduced for 'laptop' and 'bowl' category, and to low extent for the 'mug' category. Furthermore, estimating a 6D pose of non-symmetrical objects is still a challenging task, which could potentially be a good topic for future work.