Item type |
SIG Technical Reports(1) |
公開日 |
2022-11-11 |
タイトル |
|
|
タイトル |
3D Face Reconstruction-based Augmentation for Gaze and Head Redirection |
タイトル |
|
|
言語 |
en |
|
タイトル |
3D Face Reconstruction-based Augmentation for Gaze and Head Redirection |
言語 |
|
|
言語 |
eng |
資源タイプ |
|
|
資源タイプ識別子 |
http://purl.org/coar/resource_type/c_18gh |
|
資源タイプ |
technical report |
著者所属 |
|
|
|
The University of Tokyo/CyberAgent, Inc. |
著者所属 |
|
|
|
CyberAgent, Inc. |
著者所属(英) |
|
|
|
en |
|
|
The University of Tokyo / CyberAgent, Inc. |
著者所属(英) |
|
|
|
en |
|
|
CyberAgent, Inc. |
著者名 |
Jiawei, Qin
Xueting, Wang
|
著者名(英) |
Jiawei, Qin
Xueting, Wang
|
論文抄録 |
|
|
内容記述タイプ |
Other |
|
内容記述 |
Gaze and head redirection is to change the gaze and head orientation of a full-face image to a target direction. Instead of only changing face direction in existing generation tasks, controllable both gaze and head redirection containing more complicated information is also critical for creating expressive face image. However, the redirection precision of existing gaze and head redirection models critically degrades once the target direction goes out of the model's capable range, which is limited by the angle range of the training data. In this paper we propose to use monocular 3D face reconstruction as data augmentation to extend the redirection range of the existing limited real data. The augmentation can largely extend the head pose range by rotation while preserving the original gaze information of real data. Consequently, the range of the head pose and gaze can both be extended. Experiments show that the proposed data augmentation significantly improved the redirection performance especially when redirecting to a relatively large angle while keeping the image quality. |
論文抄録(英) |
|
|
内容記述タイプ |
Other |
|
内容記述 |
Gaze and head redirection is to change the gaze and head orientation of a full-face image to a target direction. Instead of only changing face direction in existing generation tasks, controllable both gaze and head redirection containing more complicated information is also critical for creating expressive face image. However, the redirection precision of existing gaze and head redirection models critically degrades once the target direction goes out of the model's capable range, which is limited by the angle range of the training data. In this paper we propose to use monocular 3D face reconstruction as data augmentation to extend the redirection range of the existing limited real data. The augmentation can largely extend the head pose range by rotation while preserving the original gaze information of real data. Consequently, the range of the head pose and gaze can both be extended. Experiments show that the proposed data augmentation significantly improved the redirection performance especially when redirecting to a relatively large angle while keeping the image quality. |
書誌レコードID |
|
|
収録物識別子タイプ |
NCID |
|
収録物識別子 |
AA11131797 |
書誌情報 |
研究報告コンピュータビジョンとイメージメディア(CVIM)
巻 2022-CVIM-231,
号 1,
p. 1-6,
発行日 2022-11-11
|
ISSN |
|
|
収録物識別子タイプ |
ISSN |
|
収録物識別子 |
2188-8701 |
Notice |
|
|
|
SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc. |
出版者 |
|
|
言語 |
ja |
|
出版者 |
情報処理学会 |