US 12,141,996 B2
Gaze estimation cross-scene adaptation method and device based on outlier guidance
Feng Lu, Beijing (CN); and Yunfei Liu, Beijing (CN)
Assigned to Beihang University, Beijing (CN)
Filed by Beihang University, Beijing (CN)
Filed on Dec. 24, 2021, as Appl. No. 17/561,866.
Claims priority of application No. 202110689959.0 (CN), filed on Jun. 22, 2021.
Prior Publication US 2022/0405953 A1, Dec. 22, 2022
Int. Cl. G06T 7/70 (2017.01); G06F 17/18 (2006.01); G06N 3/084 (2023.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01); G06V 40/18 (2022.01)
CPC G06T 7/70 (2017.01) [G06F 17/18 (2013.01); G06N 3/084 (2013.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01); G06V 40/193 (2022.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30201 (2013.01)] 6 Claims
OG exemplary drawing
 
1. A method for gaze estimation cross-scene adaptation based on outlier guidance, comprising:
performing pre-training on a source domain based on a given arbitrary gaze estimation model, to obtain a collaborative learning model group;
determining an average collaborative model corresponding to each collaborative learning model in the collaborative learning model group, to obtain an average collaborative model group;
generating an outlier corresponding to the collaborative learning model group, based on a target image, the collaborative learning model group and the average collaborative model group;
using an outlier loss function and the outlier to optimize the collaborative learning model group; and
using any collaborative learning model in the optimized collaborative learning model group to perform gaze estimation,
wherein the determining the average collaborative model corresponding to each collaborative learning model in the collaborative learning model group includes: employing a manner of exponential moving average, to determine the average collaborative model corresponding to each collaborative learning model in the collaborative learning model group.