3D object recognition from cluttered and occluded scenes with a compact local feature

This paper presents a novel local surface descriptor for 3D object recognition in the presence of clutter and occlusion. For a keypoint and its local neighbor points, a unique and repeatable local reference frame is first constructed. Then the depth information of those local points is encoded into a vector using multi-view method. Finally we obtain a compact feature with length of 27. Besides its compactness in length, the generated feature is tested to be not only descriptive on various datasets but also robust against Gaussian noise, short noise and varying mesh resolutions. Based on this descriptor, we propose a hierarchical algorithm for automatic object recognition. By feature matching, scene-point-to-model-point correspondences are established after potential planar clutters are removed from the scene. These correspondences are then utilized for object identification and pose estimation. Experimental results on three popular datasets demonstrate the effectiveness of our proposed technique which achieves the best results compared to the state of the art with all targets correctly recognized and their poses estimated.