An improved approach for model-based detection and pose estimation of texture-less objects

Abstract

Detection and pose estimation of texture-less objects still faces several challenges such as foreground occlusions, background clutter, multi-instance objects, large scale and pose changes to name but a few. In this paper, we present an improved approach for model based detection and pose estimation of texture-less objects, LINEMOD [4], in order to improve the robustness of pose estimation with partial foreground occlusions. For template creation, we modify Gradient Response Maps and propose Gradient Orientation Maps, where Non-Maximum Suppression and Dual Threshold Algorithm are applied. And we adopt image pyramid searching method for fast template matching. Next, the approximate object pose associated with each detected template is used as a starting point for fine pose estimation with Iterative Closest Point algorithm. Thirdly, we improve the accuracy of fine pose estimation by using point cloud filter. Experimental results show that our approach is more robust to estimate the pose of texture-less objects with partial foreground occlusions.