Reproduce “GeoNet: Benchmarking Unsupervised Adaptation across Geographies” on OpenResearch

About this paper

Tarun Kalluri, Wangdong Xu, Manmohan Chandraker

UC San Diego researchers introduce GeoNet, a large-scale benchmark designed to measure unsupervised adaptation across diverse global geographies. This dataset reveals significant performance declines in contemporary computer vision models and state-of-the-art domain adaptation methods when confronted with complex geographic distribution shifts.

Key insights

• Geographic domain shifts are complex and multifaceted, primarily driven by distinct context, design, and prior variations, which existing domain adaptation techniques struggle to address.
• Current state-of-the-art unsupervised domain adaptation methods largely fail to effectively bridge the geographic domain gap, often performing no better than source-only baselines and sometimes demonstrating negative transfer.
• Even large-scale pre-training on massive datasets and advanced architectures like Vision Transformers do not inherently provide sufficient robustness to geographic distribution shifts, leaving a substantial performance gap compared to target-supervised models.

Summary by alphaXiv.