US 12,080,056 B1
Performing explanation jobs for computer vision tasks
Ashish Rajendra Rathi, San Jose, CA (US); Michele Donini, Berlin (DE); Tyler Stephen Hill, Los Altos, CA (US); Krishnaram Kenthapadi, Sunnyvale, CA (US); Xinyu Liu, Santa Clara, CA (US); Pinar Altin Yilmaz, Palo Alto, CA (US); and Muhammad Bilal Zafar, Berlin (DE)
Assigned to Amazon Technologies, Inc., Seattle, WA (US)
Filed by Amazon Technologies, Inc., Seattle, WA (US)
Filed on Nov. 26, 2021, as Appl. No. 17/535,909.
Int. Cl. G06V 10/70 (2022.01); G06T 7/10 (2017.01); G06V 10/764 (2022.01); G06V 10/77 (2022.01); G06V 20/50 (2022.01)
CPC G06V 10/87 (2022.01) [G06T 7/10 (2017.01); G06V 10/764 (2022.01); G06V 10/768 (2022.01); G06V 10/7715 (2022.01); G06V 20/50 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A system, comprising:
at least one processor; and
a memory, storing program instructions that when executed by the at least one processor, cause the at least one processor to implement:
receive, via an interface for a machine learning service, a request to execute an explanation job for a computer vision machine learning model hosted by the machine learning service with respect to an image that specifies an image segmentation technique to apply to the image as part of performing the explanation job;
launch an explanation processing container hosted by the machine learning service to perform the explanation job for the computer vision machine learning model, wherein the explanation job extracts different super pixels according to the image segmentation technique from the image according to a segmentation technique to determine respective relative importance values of the different super pixels on one or more inferences generated by the computer vision machine learning model on the image and generates, based on the respective relative importance values of the different super pixels, one or more heat maps indicating the relative importance values of the different super pixels on performance of the computer vision machine learning model; and
provide, according to a specification of the explanation job, the one or more heat maps as part of a result for the explanation job.