TY - GEN
T1 - SGFNeRF
T2 - 7th Asian Conference on Pattern Recognition, ACPR 2023
AU - Zhou, Peizhu
AU - Liu, Xuhui
AU - Zhang, Baochang
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.
PY - 2023
Y1 - 2023
N2 - NeRF (Neural Radiance Fields), as an implicit 3D representation, has demonstrated the capability to generate highly realistic and dynamically consistent images. However, its hierarchical sampling approach introduces a significant amount of redundant computation, leading to erroneous geometric information, particularly in high-frequency facial details. In this paper, we propose SGFNeRF, a novel 3D face generation model by integrating a 2D CNN-based generator and face depth priors optimization method in the same framework. We employ a Gaussian distribution for sampling to extract facial surface information. Additionally, we design a feature decoder to incorporates depth uncertainty into our method, enabling the method to explore regions further away from face surfaces while preserving its ability to capture fine-grained details. We conduct experiments on the FFHQ dataset to evaluate the performance of our proposed method. The results demonstrate a significant improvement compared to previous approaches in terms of various evaluation metrics.
AB - NeRF (Neural Radiance Fields), as an implicit 3D representation, has demonstrated the capability to generate highly realistic and dynamically consistent images. However, its hierarchical sampling approach introduces a significant amount of redundant computation, leading to erroneous geometric information, particularly in high-frequency facial details. In this paper, we propose SGFNeRF, a novel 3D face generation model by integrating a 2D CNN-based generator and face depth priors optimization method in the same framework. We employ a Gaussian distribution for sampling to extract facial surface information. Additionally, we design a feature decoder to incorporates depth uncertainty into our method, enabling the method to explore regions further away from face surfaces while preserving its ability to capture fine-grained details. We conduct experiments on the FFHQ dataset to evaluate the performance of our proposed method. The results demonstrate a significant improvement compared to previous approaches in terms of various evaluation metrics.
KW - 3D scene representation
KW - Face generation
KW - Generative adversarial network
KW - Neural radiance fields
UR - https://www.scopus.com/pages/publications/85177465495
U2 - 10.1007/978-3-031-47665-5_20
DO - 10.1007/978-3-031-47665-5_20
M3 - 会议稿件
AN - SCOPUS:85177465495
SN - 9783031476648
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 238
EP - 249
BT - Pattern Recognition - 7th Asian Conference, ACPR 2023, Proceedings
A2 - Lu, Huimin
A2 - Blumenstein, Michael
A2 - Cho, Sung-Bae
A2 - Liu, Cheng-Lin
A2 - Yagi, Yasushi
A2 - Kamiya, Tohru
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 5 November 2023 through 8 November 2023
ER -