Reproduce “AffectNet+: A Database for Enhancing Facial Expression Recognition with Soft-Labels” 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

Ali Pourramezan Fard, Mohammad Mehdi Hosseini, Timothy D. Sweeny, Mohammad H. Mahoor

AffectNet+ introduces a large-scale facial expression recognition database featuring "soft-labels" that assign probability vectors across multiple emotions, addressing the ambiguity of traditional single-label datasets. The enhanced dataset, developed by the University of Denver, includes comprehensive metadata and categorizes images by recognition difficulty, providing a more nuanced foundation for training robust FER models.

Key insights

  • Soft-labels offer a more realistic and comprehensive representation of human emotional expressions than hard-labels, better capturing nuances, ambiguity, and co-occurring emotions.
  • Leveraging a multi-annotated, high-agreement subset of data (MAS) enables reliable automatic generation of soft-labels for a much larger dataset.
  • Categorizing dataset images into "Easy," "Challenging," and "Difficult" subsets, based on soft-label analysis, allows for more targeted model development and evaluation on varying levels of expression complexity.

Summary by alphaXiv.

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