GeoDiffusion: Creating Synthetic Data for Upscaling Geographically Diverse Image Datasets

Abstract

Large scale image datasets used to train leading computer vision models often contain biases in their geographic composition, often sourcing most images from North America and Western Europe, while underrepresenting regions like Africa and Asia. Attempts to address this geographical diversity bias have been made through manually curating "fair" datasets, but these datasets face problems such as their small size and expensive gathering costs. Recent advances in generative visual models like diffusion models present opportunities to efficiently generate synthetic data that can address these problems of size and resource cost for creating fair datasets. This work looks to create a pipeline for generating synthetic data for upscaling geographically diverse image datasets, and evaluates the effects of model accuracy when such synthetic data is introduced during training. We find that within our experimental setup, introducing moderate amounts of synthetic data during training increases model test accuracy for most geographic regions, but hinders test accuracy when introducing large amounts of synthetic data. We find that our synthetic data pipeline serves as an effective tool for generative data augmentation and image perturbation for geographically diverse image data. Additionally, we analyze trends in synthetic data output across image class and region, and point to further directions for investigating the use of generative visual models for creating fair synthetic data.