Most modern computer vision models share a common, invisible "worldview." Because they are trained on massive datasets scraped from the internet, they inherit the geographic distribution of the web—which is overwhelmingly concentrated in the United States and Europe. When these models are deployed in the Global South, they encounter objects, environments, and visual contexts they were never designed to understand.
The Hidden Skew of Web-Scraped Data
The performance of a vision model depends on how well its training data represents the real world. However, as Maxime Oquab and colleagues point out in DINOv2: Learning Robust Visual Features without Supervision, even "foundation" models trained on hundreds of millions of images suffer from geographic imbalance.
This isn't just a matter of having fewer photos from certain places; it is a fundamental difference in how objects appear. A "stove" in a high-income Western household looks nothing like a stove in a rural village in India. When models are trained primarily on the former, they fail to recognize the latter. Vikram V. Ramaswamy and the team behind GeoDE: a Geographically Diverse Evaluation Dataset for Object Recognition quantified this gap, finding that models often see a 7% drop in accuracy when moving from European to African contexts.
Why Scaling and Adaptation Fail
A common assumption in machine learning is that we can fix bias by either "scaling up" (training on more data) or "adapting" (tweaking a model for a new environment). Recent research suggests neither is a silver bullet for geographic shifts.
In GeoNet: Benchmarking Unsupervised Adaptation across Geographies, Tarun Kalluri demonstrates that even pre-training on billions of images fails to bridge the performance gap between regions like the USA and Asia. Similarly, Viraj Prabhu and Nikhil Naik explored Unsupervised Domain Adaptation (UDA) in Can domain adaptation make object recognition work for everyone?. They found that standard UDA techniques—which usually work well for simple shifts like "sunny to rainy"—fall short here. Geographic shifts are more complex because they involve changes in object appearance, category varieties, and the surrounding background all at once.
Beyond Recognition: Video and Quality
Geographic bias extends beyond identifying objects; it affects how we measure the quality of digital experiences. Most video quality metrics are tuned for professional content viewed on high-end hardware in the West.
To address this, Sandeep Mishra and Alan Bovik introduced the LIVE-ShareChat IUGC-VQA Database in Subjective and Objective Analysis of Indian Social Media Video Quality. By focusing on user-generated content in emerging economies, they highlight that "quality" itself is perceived differently depending on the local visual culture and the typical hardware used in that region.
Building a Representative Future
The consensus across these works is that we cannot "scrape our way" out of geographic bias. Relying on the internet as it exists today only reinforces existing imbalances.
Instead, researchers are moving toward intentional data collection. By commissioning original photos from a global workforce—as seen in the GeoDE project—and developing benchmarks like GeoNet that explicitly test for regional robustness, the field is beginning to build vision systems that work for everyone, regardless of their latitude or longitude.