Comparing RGBD-based 6D Pose Estimation

In this work the suitability of the pose estimation method "PVN3D", by He et al.[3], is evaluated for industrial applications. It first gives an overview over the research eld of 6D pose estimation. Starting by explaining the basics of the eld this work goes over point pair features, machine learning and evaluation metrics as a background knowledge for the 6D pose estimation methods of He et al.[3] and Vidal et al.[2, 23] and the 6D pose estimation benchmark "BOP" introduced by Hodan et al. [6, 16]. The main contribution of this thesis is the evaluation of PVN3D[3] on the BOP[6] framework, so it can be compared to a multitude of other pose estimators, e.g. the methods benchmarked in the BOP 2020 challenge[16]. Here we focus on the T-LESS[11] dataset, as provided by BOP[16], it provides images of objects typical of industrial applications. We found that PVN3D[3] performs worst on T-LESS[11] compared to the BOP-results from 2020[16]. Specially it performs worse than the PPF based method by Vidal et al.[2, 23], itself ranking 6th in 2020[16]. However, it has to be noted, that our experiments were not fully conclusive, as there are strong indications that there is an unknown bug in the training of PVN3D[3].