Reproduce “Training Deep Networks for Facial Expression Recognition with Crowd-Sourced Label Distribution” 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

Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, Zhengyou Zhang

Researchers at Microsoft Research introduced the FER+ dataset, re-annotating the FER dataset with crowd-sourced label distributions derived from 10 taggers per image. Their work demonstrates that training deep convolutional neural networks for facial expression recognition using these label distributions, specifically with Probabilistic Label Drawing or Cross-Entropy Loss, yields an accuracy improvement exceeding 1% compared to conventional majority voting methods.

Key insights

  • Leveraging the full crowd-sourced label distribution directly, instead of relying on a single majority vote, significantly enhances the accuracy of deep networks for facial expression recognition.
  • Probabilistic Label Drawing and using Cross-Entropy Loss with the entire label distribution are effective strategies for handling inherent noise and subjectivity in crowd-sourced annotations during DCNN training.
  • The quality of crowd-sourced labels for subjective tasks markedly improves with an increased number of annotators per item, underscoring the value of collecting label distributions.

Summary by alphaXiv.

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