Reproduce “Regional biases in image geolocation estimation: a case study with the SenseCity Africa dataset” 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
Ximena Salgado Uribe, Martí Bosch, Jérôme Chenal
An analysis of a leading image geolocation model on an African dataset reveals significant regional biases, with accuracy substantially reduced and predictions skewed towards higher-income Western regions, underscoring the impact of imbalanced training data.
Key insights
- The ISNs model exhibited significantly lower accuracy (e.g., 1.0% at street level for SCA100 vs. 9.7% for IM2GPS3k) when applied to images from the African continent.
- The model displayed a strong tendency to mispredict image locations towards higher-income groups and Western regions, even for images originating from low-income African contexts.
- The magnitude of prediction error correlated with an increased bias towards mispredicting locations in higher-income groups and Western regions across clustered error categories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.