Brief

Researchers have developed MAPRPose, a novel two-stage framework for estimating the 6D pose of multiple objects in challenging, cluttered environments. The system first generates pose hypotheses using mask-aware correspondences and then refines these using an amodal mask prediction and region-of-interest re-alignment module. This approach significantly improves accuracy and speed, achieving state-of-the-art performance on the BOP benchmark. AI

Researchers have developed a new method for 3D-to-2D liver registration in laparoscopic surgery, aiming to improve augmented reality (AR) guidance. This approach eliminates the need for complex finite-element models by combining robust rigid initialization with patient-specific non-rigid refinement. The system utilizes depth augmentation and contour maps for initial alignment and a statistical deformation model for refinement, achieving a mean target registration error of 14.73 mm on a clinical dataset. AI

Researchers have developed an ensemble directional Kalman filter (EnDKF) for improved pose tracking. This method integrates unit-quaternions to better represent directional uncertainty, moving beyond traditional Kalman filter assumptions. Experiments using the FoundationPose algorithm on a head-tracking scenario showed a significant reduction in error compared to using measurements alone. AI