Erratum to "Multi-View Face Synthesis via Progressive Face Flow"

Yangyang Xu, Xuemiao Xu, Jianbo Jiao, Keke Li, Cheng Xu, Shengfeng He

Abstract

This erratum serves as a formal correction to the paper "Multi-View Face Synthesis via Progressive Face Flow," originally published in IEEE Transactions on Image Processing in 2021. The primary purpose of this document is to restore Figure 5, which was incorrectly rendered with missing images in the initial publication. This figure provides the essential qualitative evidence required to evaluate the performance of the proposed synthesis method across a wide range of facial poses.

Introduction

Multi-view face synthesis is the task of generating a realistic image of a person’s face at a specific target angle (e.g., a profile view) based on an input image from a different angle (e.g., a frontal view). The original 2021 research introduced a novel approach to this problem. However, a technical publishing error resulted in an incomplete Figure 5, preventing readers from visually assessing the model's output. This erratum restores the missing visual data to complete the scientific record.

Methodology and Dataset

The core technique presented in the original work is Progressive Face Flow. This method likely utilizes a multi-stage refinement process to incrementally "flow" or warp facial features from a source pose to a target pose, ensuring that identity and texture consistency are maintained even across large angular gaps. To evaluate this method, the authors utilized the Multi-PIE dataset, a standard benchmark for facial recognition and synthesis that includes variations in pose, illumination, and expression.

Qualitative Results

The restored Figure 5 provides a comprehensive qualitative comparison of the model's performance. The figure displays synthesis results at five distinct angles: 0°, 15°, 45°, 60°, and 90°.

Qualitative results of the Progressive Face Flow method on the Multi-PIE dataset. Odd columns show synthesized results; even columns show ground truth.

Key insights derived from the restored figure include:

  • Visual Consistency: The model successfully maintains high-fidelity facial features, such as eyes, facial hair, and glasses, even when rotating a face from a frontal (0°) to a profile (90°) view.
  • Feature Hallucination: The results highlight the model's ability to accurately hallucinate missing information—such as the side of a nose or an ear—that is not visible in the source input.
  • Robustness to Occlusion: The inclusion of subjects wearing glasses demonstrates that the "Face Flow" method can handle complex occlusions and reflections during the synthesis process.
  • Ground Truth Fidelity: The synthesized images (odd columns) appear remarkably similar to the ground truth images (even columns) in terms of lighting, skin tone, and structural identity.

Conclusion

The restoration of Figure 5 confirms the qualitative success of the Progressive Face Flow method. By providing a side-by-side comparison with ground truth data across a 90-degree range, this erratum validates the model's ability to generate realistic, identity-preserving facial images from varying viewpoints.

Summary updated

Abstract

In the above article <xref ref-type="bibr" rid="ref1" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[1]</xref> , unfortunately, <xref ref-type="fig" rid="fig1" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Fig. 5</xref> was not displayed correctly with many empty images. The correct version is supplemented here.
BibTeX
@article{xu2021erratum,
  author = {Yangyang Xu and Xuemiao Xu and Jianbo Jiao and Keke Li and Cheng Xu and Shengfeng He},
  title = {{Erratum to "Multi-View Face Synthesis via Progressive Face Flow"}},
  journal = {IEEE Trans. Image Process.},
  year = {2021},
  doi = {10.1109/tip.2021.3097165},
}
Figure