US 11,055,572 B2
System and method of training an appearance signature extractor
Hiroyuki Uchiyama, Kanagawa-ken (JP)
Assigned to Canon Kabushiki Kaisha, Tokyo (JP)
Filed by CANON KABUSHIKI KAISHA, Tokyo (JP)
Filed on Feb. 22, 2019, as Appl. No. 16/283,583.
Prior Publication US 2020/0272860 A1, Aug. 27, 2020
Int. Cl. G06K 9/62 (2006.01); G06N 20/00 (2019.01); G06K 9/03 (2006.01)
CPC G06K 9/6262 (2013.01) [G06K 9/03 (2013.01); G06K 9/6257 (2013.01); G06N 20/00 (2019.01)] 10 Claims
OG exemplary drawing
 
1. A method of training an appearance signature extractor using a training data set, the training data set including input images associated with a plurality of domains, the method comprising:
inputting an input image of the training data set, an identity label, and a domain label to the appearance signature extractor, wherein the identity label provides identity information of the input image, and the domain label provides domain information of the input image, the domain information corresponding to one of the plurality of domains;
determining an appearance signature from the input image using the appearance signature extractor;
determining identity scores for the plurality of domains respectively, based on the appearance signature;
determining a loss value from among loss values respectively calculated for the plurality of domains according to the domain label, wherein the loss values relate to an error between the identity score label and the identity scores; and
updating the appearance signature extractor using the determined loss value.