EE ZOOM Seminar: Best Buddies Registration For Point Clouds
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בצ'אט
Join Zoom Meeting
https://zoom.us/j/97758431884?pwd=eDYxT1ZGdk1BSHZ4cjdhbE0xREk0UT09
Meeting ID: 977 5843 1884
Password: 843294
Speaker: Tal Dimry
M.Sc. student under the supervision of Prof. Shai Avidan and Dr. Raja Giryes
Wednesday, May 13th, 2020 at 11:00 AM
Best Buddies Registration For Point Clouds
Abstract
We propose a new loss function for the point clouds registration problem. The loss is based on the Best Buddies Similarity (BBS) measure that counts the number of mutual nearest neighbors between two point clouds. This measure has been shown to be robust to outliers and missing data in the case of template matching for images. BBS is not differentiable, because it involves a nearest neighbor search, and we overcome this by using a differentiable approximation that is based on the soft argmin operator. This allows us to perform gradient descent using existing neural network optimization tools.
We explore two variants of our loss function. One with a very large basin of convergence and one with a narrower basin but with high accuracy. Combining the two leads to an efficient and accurate registration of two point clouds without requiring any pre- or post-registration steps. We also show that our loss can use the normals of the points, if they are available, to further improve accuracy. Exper- iments on various data sets, both synthetic and real, demonstrate the effectiveness of our approach.
