Instance- and Category-Level 6D Object Pose Estimation

Abstract

Interest in estimating the 6D pose, i.e. 3D locations and rotations, of an object of interest has emerged since its promising applications in fields such as robotics and augmented reality. To recover poses from objects that have been seen in advance, instance-level methods have been presented to overcome challenges such as occlusion, clutter and similarly looking distractors. The problem has recently been addressed at the category level, where poses of object instances from a given category that have not been seen a priori are estimated, introducing new challenges such as distribution shifts and intra-class variations. In this chapter, the 6D object pose estimation problem at the levels of both instances and categories is presented, discussed, and analysed by following the available literature on the topic. First, the problem and its associated challenges are formulated and presented. To continue, instance-level methods are dissected depending on their architectures and category-level methods are examined according to their search space dimension. Popular datasets, benchmarks and evaluation metrics on the problem are presented and studied with respect to the challenges that they present. Quantitative results of experiments available in the literature are analysed to determine how methods perform when presented with different challenges. The analyses are further extended to compare three methods by using our own implementations aiming to solidify already published results. To conclude, the current state of the field is summarised and potential future research directions are identified.