3D Correspondences by Local Feature Matching

Guy Berdugo,  Guy Gilboa

Electrical Engeering Deapartment,

Technion Israel Institute of technology, Haifa, Israel 

3D Matches Example
3D Matches Example


There is a growing trend today of incorporating 3D sensors within mobile devices. The heavy power limitations naturally lead to a considerable decrease in the quality of point cloud data. This requires highly robust algorithms in order to produce reliable descriptors. 3D descriptors are essential in solving problems such as point cloud registration, object reconstruction, recognition and tracking.
In this work we propose a new local descriptor, Histogram of Relative Angles (HoRA), which is designed to perform robustly for various data degradations of noise and artifacts. The method is simple in its logic and ecient computationally. The descriptor is based on two steps, computation of a local reference frame (LRF) and relative angles. We propose an improvement to existing state-of-the-art LRF computation algorithm which yields a notable improvement in it repeatability rate for various degradations. The relative angles are shown to be highly descriptive geometric features, yet more robust than surface normal estimation or point distribution, commonly used in most recent state-of-the-art methods (SHOT, USC, Spin Image, ThrIFT, RoPS). We show that HoRA can cope well with noise degradation and low resolution e ects. Comprehensive experiments on various benchmarks demonstrate the clear superiority of this approach for moderate-to-low quality input over numerous state of the art algorithms


HoRA feature extraction illustration


Algorithm’s flowchart: 3D surface matching



Matlab Code ,including 3D CV Lab dataset


Contact us at: