For those that are working with 3D point clouds (for example, 3D keypoint detection, 3D local feature descriptors, 3D object recognition/classification/retrieval, 3D scene modeling and reconstruction, and 3D data registration, among many other possible applications), this is a must-read scientific paper. Also, as indicated by the authors, this paper “serve[s] as a user guide for the selection of the most appropriate feature descriptor in the area of 3D computer vision.”
The authors present an extensive review of different 3D feature extraction techniques (associated with a wide list of relevant related publications), focusing the performance evaluation on these 10 methods: SI, 3DSC, LSP, THRIFT, PFH, FPFH, SHOT, USC, RoPS, and TriSI (implementations in Point Cloud Library and MATLAB). They present an extensive comparison of the previously cited techniques considering: (1) their robustness to noise disturbances (different types of noise); (2) variations of mesh resolution; (3) distance to the mesh boundary; (4) keypoint localization error; (5) occlusion and clutter; (6) scalability (small, medium, or large number of points); (7) diverse acquisition techniques and devices (for example, LIDAR, Stereo, Kinect); (8) different application scenarios; (9) descriptiveness of a feature descriptor; and (10) computational efficiency. The evaluation was performed using eight publicly available data sets, each one with different characteristics, allowing the authors to demonstrate which method is better for which situation/application, and thus helping in the selection of the most appropriate descriptor according to the user’s needs.