US 11,875,527 B1
Descriptor generation and point cloud fusion
Lina M. Paz-Perez, San Jose, CA (US); Chavdar Papazov, Sunnyvale, CA (US); Vimal Thilak, Daly City, CA (US); and Jai Prakash, Santa Clara, CA (US)
Assigned to Apple Inc., Cupertino, CA (US)
Filed by Apple Inc., Cupertino, CA (US)
Filed on Apr. 7, 2021, as Appl. No. 17/224,192.
Claims priority of provisional application 63/008,929, filed on Apr. 13, 2020.
Int. Cl. G06T 7/73 (2017.01); G06T 7/11 (2017.01); G06T 7/64 (2017.01); G06T 17/00 (2006.01); G06F 18/25 (2023.01)
CPC G06T 7/73 (2017.01) [G06F 18/253 (2023.01); G06T 7/11 (2017.01); G06T 7/64 (2017.01); G06T 17/00 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/20132 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A method comprising:
at an electronic device having a processor:
obtaining an image of a physical environment;
determining a subset of points of the image and orientations of the subset of points using a neural network comprising layers, the layers comprising a precomputed-weight layer configured to compute an orientation for each point of the subset of points based on:
identifying an image patch to associate with each point of the subset of points; and
using the image patch associated with each point of the subset of points to determine the orientation for each point of the subset of points, wherein the orientation of each point of the subset of points is determined formulaically via the precomputed-weight layer, wherein the orientation determined by the precomputed-weight layer is independent of weights of the layers that were learned during training;
selecting crops of the image for the subset of points based on the orientations; and
generating descriptors for the subset of points of the image based on the crops of the image.