Reproduce “PoseFace: Pose-Invariant Features and Pose-Adaptive Loss for Face Recognition” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Qiang Meng, Xiaqing Xu, Xiaobo Wang, Yang Qian, Yunxiao Qin, Zezheng Wang, Chenxu Zhao, Feng Zhou and 1 more

PoseFace employs a self-disentanglement method to project facial identity and pose features into orthogonal subspaces, coupled with a pose-adaptive loss to mitigate data imbalance between frontal and profile views. The framework achieved enhanced face recognition accuracy, particularly for extreme pose variations, across multiple benchmarks while maintaining computational efficiency.

Key insights

  • The effective disentanglement of identity and pose information into orthogonal subspaces is critical for improving face recognition accuracy under extreme pose variations.
  • Leveraging "pure" pose features derived from facial landmarks, rather than direct regression, is essential for robust disentanglement, preventing identity information from contaminating the pose representation.
  • Dynamically adjusting loss margins based on facial pose angles enables the model to focus training on difficult-to-recognize profile faces, thereby counteracting data imbalance and enhancing overall robustness.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.