Sign in to view source links and access this dataset
Description
Face4FairShifts is a large benchmark dataset containing 100,000 original face images across four visual domains, including 30,000 in the Photo domain and 25,000 each in Art and other domains. It was created by Tianjin University and is hosted on Hugging Face. The dataset was last updated on May 19, 2025.
Use Cases
Benchmarking fairness algorithms based on the multi-domain face image collection
Training models for domain adaptation based on images from Photo, Art, and other visual domains
Evaluating algorithmic bias and robustness across different visual representations of faces
Conducting research on demographic parity and shift in computer vision systems
Strengths
Contains 100,000 original face images, providing a substantial scale for training and evaluation
Explicitly structured across four visual domains, including 30,000 Photo and 25,000 Art images
Designed specifically as a benchmark for fairness and robust learning, providing a clear research focus
Limitations
Column-level documentation is absent; field semantics must be inferred after download
Description metadata is limited; actual data quality requires manual inspection after download
Provenance
Source
Tianjin University
Collection Method
Likely collected or curated from various visual sources representing different domains.
Time Range
null
Freshness
Last updated 2025-05-19 15:14:26; freshness should be verified
Geography
null
Use of this dataset is restricted to non-commercial research and educational purposes.