US 12,468,939 B2
Object discovery using an autoencoder
Hankyu Moon, San Ramon, CA (US); Heng Hao, San Jose, CA (US); Sima Didari, San Jose, CA (US); Jae Oh Woo, Fremont, CA (US); and Patrick David Bangert, Sunnyvale, CA (US)
Assigned to SAMSUNG SDS AMERICA, INC., San Jose, CA (US)
Filed by Samsung SDS America, Inc., San Jose, CA (US)
Filed on Nov. 2, 2021, as Appl. No. 17/517,313.
Claims priority of provisional application 63/193,972, filed on May 27, 2021.
Prior Publication US 2022/0383105 A1, Dec. 1, 2022
Int. Cl. G06N 3/08 (2023.01); G06F 18/2113 (2023.01); G06F 18/214 (2023.01); G06F 18/22 (2023.01)
CPC G06N 3/08 (2013.01) [G06F 18/2113 (2023.01); G06F 18/2155 (2023.01); G06F 18/22 (2023.01)] 22 Claims
OG exemplary drawing
 
1. A method comprising:
discovering, using a modulated contrastive loss, a first object in a plurality of training images, wherein the plurality of training images includes a first training image;
embedding a description of the first object in a pattern space, wherein the pattern space is a subset of a latent space of an autoencoder;
identifying, using the pattern space, a second object in a first data image, wherein a second plurality of images includes the first data image; and
presenting a first annotated image of the first data image on a display screen, wherein a first bounding box of the first annotated image is associated with a first pattern vector in the pattern space,
wherein the modulated contrastive loss is in the form of an objectness score multiplied by a contrastive loss value, the objectness score is based on a combination of a histogram score with a background score, the background score is a distance from the first pattern vector to a background cluster center, and the background cluster center is found using a K-means algorithm applied to pairs of patches from the plurality of training images.