docs: fill in dataset card sections (bias, limitations, curation rationale)

#6

Fills in the dataset card sections that were previously marked as [More Information Needed].

Changes:

  • Curation Rationale: describes why the 101 categories were chosen and the benchmark's design goals
  • Source Data: documents that images came from Foodspotting.com and describes the data collection process
  • Annotations: clarifies the difference between cleaned test labels and noisy training labels
  • Personal and Sensitive Information: notes restaurant-setting photography context
  • Social Impact: describes deployment contexts (dietary apps, nutrition tracking, recommendation systems) and global coverage limitations
  • Discussion of Biases: documents category selection bias (Western restaurant cuisine skew), dietary category imbalance (~12-15% plant-based classes), label noise in training split, and photography style bias
  • Other Known Limitations: covers noisy training labels, test set statistical power, missing nutritional metadata, and points to extended benchmarks with broader international coverage
  • Dataset Curators: identifies ETH Zurich authors (Bossard, Guillaumin, Van Gool, ECCV 2014)

All information is sourced from the original paper and the dataset homepage.

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment