Detection and fine 3D pose estimation of texture-less objects

Abstract

This paper presents a method for real-time detection and accurate estimation pose of the object. We start from the LINEMOD which be proposed by Hinterstoisser et al. However, it show typical problems such as not being robust to clutter and occlusions. In this paper, we propose a method, namely Block-Matching. Firstly, extracting multi-angle templates for each object. Then cutting each template into many small blocks, and treat each block as a template to match the object in the scene. If the matching score of a small block in a template exceeds a certain threshold, it indicates that there is an object which we want in the scene. The ICP algorithm is the most common choice for fine 3D pose estimation. But ICP can easily get stuck in local minima and its performance largely depends on the quality of initialization. Particle swarm optimization is employed to refine the 6D pose of the target object in this paper to avoid these defects. Extensive experimental results demonstrate the superior performance of the approach compared to the state of the art.