Fused Pose Estimation Using Geometric and Texture Information

This paper addresses the problem of estimating the alignment pose between two models by using structure-specific local descriptors which are generated by combining 2D image data and 3D contextual shape data. The 2D texture information is represented by a robust SIFT descriptor, and the geometric information is represented by a histogram supported by the orders of curvature and angles between normal vectors. The proposed pose estimation method combines the homography relation and epipolar geometry (fundamental matrix) with the EPnP algorithm. We present experiments both in controlled and real-life scenarios to validate the approach. The quantitative evaluations show that the proposed fused pose estimation provides lower mean square registration error compared to EPnP algorithm.