The subject matter of this paper is 3D object recognition, particularly the detection of “objects of interest from cluttered 3D scenes” along with 3D poses of the objects. In fact, the authors study instance-based 3D object recognition, which is a ripe topic in the literature; however, they focus on mid-level shape representation as opposed to low-level local shape features.

The main contribution of this paper is a new framework that works in two phases. “Given the 3D point cloud of a query object” in the first phase, “candidate 3D local shapes are ... sampled from the 3D point cloud”; subsequently, through the proposed distinctive metric model, “informative and discriminative local shapes are selected,” thereby establishing the 3D shape primitives “for representing the discriminative local structures of the query object.” The second phase deals with 3D object recognition and pose estimation. Here, given the cluttered scene, the 3D local shapes are extracted to be “matched with the pre-established discriminative shape primitives,” and the geometrical information thereof are “transferred to the corresponding local shapes in the scene.”

The authors’ extensive experiments “on several popular and public 3D recognition datasets” establish the effectiveness and robustness of their methods “in the presence of noise, varying model resolutions, clutter and occlusion.” They compare their pipeline with some state-of-the-art pipelines and conclude that “mid-level discriminative shape representation can achieve superior robustness than ... low-level shape representation.” They also evaluate the scalability of their approach and claim that their pipeline is parallelizable.