US 11,727,710 B2
Weakly supervised semantic parsing
Yuncheng Li, Los Angeles, CA (US); Linjie Yang, Los Angeles, CA (US); Ning Zhang, Los Angeles, CA (US); and Zhengyuan Yang, Rochester, NY (US)
Assigned to Snap Inc., Santa Monica, CA (US)
Filed by Snap Inc., Santa Monica, CA (US)
Filed on Oct. 22, 2021, as Appl. No. 17/508,384.
Application 17/508,384 is a continuation of application No. 16/450,376, filed on Jun. 24, 2019, granted, now 11,182,603.
Claims priority of provisional application 62/725,897, filed on Aug. 31, 2018.
Prior Publication US 2022/0044010 A1, Feb. 10, 2022
This patent is subject to a terminal disclaimer.
Int. Cl. G06V 40/10 (2022.01); G06F 17/15 (2006.01); G06T 7/136 (2017.01); G06N 3/04 (2023.01)
CPC G06V 40/107 (2022.01) [G06F 17/15 (2013.01); G06N 3/04 (2013.01); G06T 7/136 (2017.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
generating an initial segmentation for a training image, the initial segmentation indicating, without body part annotations, probabilities that a portion of the training image corresponds to particular body parts of a figure, the probabilities based on joints in the figure;
training a first model based on the training image and the initial segmentation;
generating a new segmentation for the training image based on the trained first model;
refining the new segmentation using a conditional random field (CRF) model;
reinitializing the trained first model; and
re-training the first model based on the training image and the refined new segmentation.