US 12,494,046 B2
Bayesian semantic segmentation active learning with beta approximation
Sima Didari, San Jose, CA (US); Jae Oh Woo, Fremont, CA (US); Heng Hao, San Jose, CA (US); Hankyu Moon, San Ramon, CA (US); and Patrick 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 Feb. 27, 2023, as Appl. No. 18/114,820.
Claims priority of provisional application 63/341,788, filed on May 13, 2022.
Prior Publication US 2023/0368507 A1, Nov. 16, 2023
Int. Cl. G06V 10/774 (2022.01); G06N 3/047 (2023.01); G06V 10/26 (2022.01); G06V 10/764 (2022.01); G06V 10/778 (2022.01); G06V 10/82 (2022.01); G06V 20/70 (2022.01)
CPC G06V 10/7753 (2022.01) [G06N 3/047 (2023.01); G06V 10/26 (2022.01); G06V 10/764 (2022.01); G06V 10/778 (2022.01); G06V 10/82 (2022.01); G06V 20/70 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A method of training a machine vision model for recognition of scene content, the method comprising:
obtaining a first plurality of labels for a first subset of pixels in a training image, wherein the first plurality of labels are drawn from a set of class labels;
training a segmentation model based on the first plurality of labels, wherein the segmentation model is the machine vision model;
predicting a plurality of probability distributions for respective remaining pixels in the training image, wherein the remaining pixels do not have labels;
identifying, with an acquisition function, a second subset of remaining pixels with a greatest level of uncertainty, wherein the acquisition function uses a first mutual information of MJEnt and a second mutual information measure of BALD;
obtaining a second plurality of labels, from the set of class labels, for the second subset;
training the segmentation model based on the second subset using the second plurality of labels;
receiving a data image for analysis; and
performing machine vision on the data image using the segmentation model to identify a third plurality of labels, from the set of class labels, corresponding to respective regions in the data image.