Ambiguity-Aware Multi-Object Pose Optimization for Visually-Assisted Robot Manipulation

6D object pose estimation aims to infer the relative between and camera using a single image or multiple images. Most works have focused on predicting without associated uncertainty under occlusion structural ambiguity (symmetricity). However, these demand prior information about shape attributes, this condition is hardly satisfied in reality; even asymmetric objects may be symmetric viewpoint change. In addition, acquiring fusing diverse sensor data challenging when extending them robotics applications. Tackling limitations, we present an ambiguity-aware network, PrimA6D++, as generic prediction method. The major challenges estimation, such symmetry, can handled manner based measured of prediction. Specifically, devise network reconstruct three rotation axis primitive images target predict underlying along each axis. Leveraging estimated uncertainty, then optimize multi-object poses visual measurements by treating it SLAM problem. proposed method shows significant performance improvement T-LESS YCB-Video datasets. We further demonstrate real-time scene recognition capability for visually-assisted robot manipulation.