US 12,272,123 B2
Training a generator for generating realistic images using a semantically segmenting discriminator
Anna Khoreva, Stuttgart (DE); Edgar Schoenfeld, Tuebingen (DE); Vadim Sushko, Stuttgart (DE); and Dan Zhang, Leonberg (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Appl. No. 17/911,987
Filed by Robert Bosch GmbH, Stuttgart (DE)
PCT Filed Aug. 20, 2021, PCT No. PCT/EP2021/073121
§ 371(c)(1), (2) Date Sep. 15, 2022,
PCT Pub. No. WO2022/043203, PCT Pub. Date Mar. 3, 2022.
Claims priority of application No. 10 2020 210 711.4 (DE), filed on Aug. 24, 2020.
Prior Publication US 2023/0134062 A1, May 4, 2023
Int. Cl. G06N 3/08 (2023.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01); G06V 10/776 (2022.01); G06V 20/58 (2022.01); G06V 20/70 (2022.01)
CPC G06V 10/774 (2022.01) [G06V 10/764 (2022.01); G06V 10/776 (2022.01); G06V 20/58 (2022.01); G06V 20/70 (2022.01)] 11 Claims
OG exemplary drawing
 
1. A method for training a generator to generate images from a semantic map that assigns to each pixel of the generated images a semantic meaning of an object to which said each pixel belongs, the method comprising the following steps:
providing respective real training images, and associated semantic training maps that each assign a semantic meaning to each pixel of the respective training image;
generating images, using the generator, from at least one of the semantic training maps;
ascertaining at least one real training image for the at least one semantic training map;
supplying each image of the images generated by the generator and the at least one real training image that belong to the same at least one semantic training map to a discriminator, the discriminator ascertaining a semantic segmentation of said each image, the segmentation assigning a semantic meaning to each pixel of said each image;
evaluating, from each of the semantic segmentations ascertained by the discriminator, whether said each image supplied to the discriminator is a generated image or a real training image;
adjusting generator parameters that characterize a behavior of the generator with a goal that the images generated by the generator are misclassified as real images; and
adjusting discriminator parameters that characterize a behavior of the discriminator with a goal of improving an accuracy of distinguishing between generated images and real images.