US 12,462,574 B2
Method and device for evaluating an image classifier
Christian Heinzemann, Ludwigsburg (DE); Christoph Gladisch, Renningen (DE); Jens Oehlerking, Stuttgart (DE); Konrad Groh, Stuttgart (DE); Matthias Woehrle, Bietigheim-Bissingen (DE); Michael Rittel, Markgroeningen (DE); Oliver Willers, Korb (DE); and Sebastian Sudholt, Besigheim (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Appl. No. 17/790,578
Filed by Robert Bosch GmbH, Stuttgart (DE)
PCT Filed Feb. 8, 2021, PCT No. PCT/EP2021/052931
§ 371(c)(1), (2) Date Jul. 1, 2022,
PCT Pub. No. WO2021/165077, PCT Pub. Date Aug. 26, 2021.
Claims priority of application No. 10 2020 201 939.8 (DE), filed on Feb. 17, 2020.
Prior Publication US 2023/0038337 A1, Feb. 9, 2023
Int. Cl. G06V 20/58 (2022.01); G06T 1/00 (2006.01); G06V 10/25 (2022.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01); G06V 10/776 (2022.01); G06V 20/64 (2022.01)
CPC G06V 20/58 (2022.01) [G06T 1/0014 (2013.01); G06V 10/25 (2022.01); G06V 10/764 (2022.01); G06V 10/7747 (2022.01); G06V 10/776 (2022.01); G06V 20/64 (2022.01); G06V 2201/06 (2022.01)] 9 Claims
OG exemplary drawing
 
8. A training device configured to:
evaluate an image classifier, in which a classifier output of the image classifier is provided for an actuation of an at least semi-autonomous robot, the evaluating including:
ascertaining a first dataset, the first dataset including images, and annotations being assigned to the images, the annotations including information about scenes imaged in the respective images and/or about image regions to be classified and/or about movement information of the robot;
ascertaining regions of the scenes that are reachable by the robot based on the annotations;
ascertaining relevance values for image regions to be classified by the image classifier, wherein the image regions to be classified are assigned to an object in each case, and wherein the ascertaining of the relevance values includes the following steps:
ascertaining depth information of the objects;
determining a relationship between object positions and reachable regions based on the depth information; and
ascertaining the relevance values based on the relationship;
classifying the image data of the first image dataset using the image classifier; and
evaluating the image classifier based on those of the image regions correctly classified by the image classifier and incorrectly classified image
regions, and the ascertained relevance values of the image regions, the evaluating of the image classifier including determining an evaluation measure;
wherein the training device is configured to perform a follow-up training of the image classifier based on a second image dataset when the evaluation measure is worse than a predefined evaluation threshold value.