TY - GEN
T1 - Joint Geometrical and Statistical Alignment Using Triplet Loss for Deep Domain Adaptation
AU - Satya Rajendra Singh, R.
AU - Sanodiya, Rakesh Kumar
AU - Arun, P. V.
N1 - Publisher Copyright:
© 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.
PY - 2022/1/1
Y1 - 2022/1/1
N2 - Although the primitive and deep learning methods have made significant progress, problems can arise if there are large differences or distribution gaps between the training and test images. To overcome this problem, shallow and deep domain adaptation (DA) approaches have been developed. However, none of the existing deep DA approaches reduce the disparity in distribution among domains statistically and geometrically and never attempt to reduce the distance among congruent images and maximize the distance between incongruous images. Therefore, in this paper, we introduce a joint geometrical and statistical alignment using the triplet loss (JGSAT) method for deep domain adaptation. More specifically, the JGSAT reduces domain shift between the domains statistically and geometrically simultaneously by incorporating maximum mean discrepancy (MMD), CORrelation ALignment (CORAL), and triplet loss in a unified framework. Ubiquitous evaluations have affirmed that the suggested method JGSAT remarkably vanquishes cutting-edge shallow and deep domain adaptation techniques on the PIE face recognition dataset.
AB - Although the primitive and deep learning methods have made significant progress, problems can arise if there are large differences or distribution gaps between the training and test images. To overcome this problem, shallow and deep domain adaptation (DA) approaches have been developed. However, none of the existing deep DA approaches reduce the disparity in distribution among domains statistically and geometrically and never attempt to reduce the distance among congruent images and maximize the distance between incongruous images. Therefore, in this paper, we introduce a joint geometrical and statistical alignment using the triplet loss (JGSAT) method for deep domain adaptation. More specifically, the JGSAT reduces domain shift between the domains statistically and geometrically simultaneously by incorporating maximum mean discrepancy (MMD), CORrelation ALignment (CORAL), and triplet loss in a unified framework. Ubiquitous evaluations have affirmed that the suggested method JGSAT remarkably vanquishes cutting-edge shallow and deep domain adaptation techniques on the PIE face recognition dataset.
KW - Deep domain adaptation
KW - Domain adaptation
KW - Transfer learning
KW - Triplet Loss
UR - http://www.scopus.com/inward/record.url?scp=85142677708&partnerID=8YFLogxK
U2 - 10.1007/978-981-19-4453-6_8
DO - 10.1007/978-981-19-4453-6_8
M3 - Conference contribution
AN - SCOPUS:85142677708
SN - 9789811944529
T3 - Lecture Notes in Electrical Engineering
SP - 119
EP - 130
BT - Responsible Data Science - Select Proceedings of ICDSE 2021
A2 - Mathew, Jimson
A2 - Santhosh Kumar, G.
A2 - Padmanabhan, Deepak
A2 - Jose, Joemon M.
PB - Springer Science and Business Media Deutschland GmbH
T2 - 7th International Conference on Data Science and Engineering, ICDSE 2021
Y2 - 17 December 2021 through 18 December 2021
ER -