Skip to main navigation Skip to search Skip to main content

SGFNeRF: Shape Guided 3D Face Generation in Neural Radiance Fields

  • Peizhu Zhou
  • , Xuhui Liu
  • , Baochang Zhang*
  • *Corresponding author for this work
  • Beihang University

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

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.

Original languageEnglish
Title of host publicationPattern Recognition - 7th Asian Conference, ACPR 2023, Proceedings
EditorsHuimin Lu, Michael Blumenstein, Sung-Bae Cho, Cheng-Lin Liu, Yasushi Yagi, Tohru Kamiya
PublisherSpringer Science and Business Media Deutschland GmbH
Pages238-249
Number of pages12
ISBN (Print)9783031476648
DOIs
StatePublished - 2023
Event7th Asian Conference on Pattern Recognition, ACPR 2023 - Kitakyushu, Japan
Duration: 5 Nov 20238 Nov 2023

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume14408 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference7th Asian Conference on Pattern Recognition, ACPR 2023
Country/TerritoryJapan
CityKitakyushu
Period5/11/238/11/23

Keywords

  • 3D scene representation
  • Face generation
  • Generative adversarial network
  • Neural radiance fields

Fingerprint

Dive into the research topics of 'SGFNeRF: Shape Guided 3D Face Generation in Neural Radiance Fields'. Together they form a unique fingerprint.

Cite this