Object Pose Estimation

Object pose estimation determines where an object is and how it is oriented, often as a 6D pose and sometimes including its size or articulation. Methods use keypoints or dense correspondences with geometric solvers such as PnP and ICP, alongside learned RGB-D and point-cloud models, foundation features, and generative pose prediction. Symmetry-aware representations, temporal tracking, and tactile or multi-view cues help address occlusion and ambiguous geometry, with benchmarks such as BOP, LINEMOD, and NOCS testing different settings. A central challenge is making systems generalize beyond familiar objects and controlled benchmarks to cluttered, deformable, reflective, or otherwise changing real-world scenes while remaining reliable for tasks such as robotic grasping.