3D Pose Estimation and Segmentation Using Specular Cues

The authors demonstrated that specular reflections can be used for localization and pose estimation of specular objects using a known CAD model. They showed that simple feature/ intensity matching can surprisingly handle textureless and highly specular objects in challenging scenes with cluttered background, inter-reflections and partial occlusions. The authors approach uses monocular camera images and does not require 3D scanning of the target scene. The proposed approach uses simple matching cost functions and optimization algorithms, and is fast and easy to implement.