Artifacts of Idiosyncracy in Global Street View Data

TL;DR

The paper investigates biases in global street view data by shifting focus from binary road-coverage to the distribution of imagery across a city. It introduces a three-stage framework—data collection, density estimation, and distribution comparison—to quantify how actual coverage deviates from a uniform road-based prior, using KL-Divergence and Earth Mover's Distance. A large-scale analysis of 28 cities reveals substantial variation in coverage distributions, with Mapillary generally showing more uneven patterns than Google Street View and with distinct differences between global North and South contexts. A detailed Amsterdam case study, including six semi-structured interviews, shows that idiosyncrasies in deployment, driver behavior, and organizational practices drive many biases, underscoring the need for domain-aware data collection and bias-addressing strategies in urban computer vision datasets.

Abstract

Street view data is increasingly being used in computer vision applications in recent years. Machine learning datasets are collected for these applications using simple sampling techniques. These datasets are assumed to be a systematic representation of cities, especially when densely sampled. Prior works however, show that there are clear gaps in coverage, with certain cities or regions being covered poorly or not at all. Here we demonstrate that a cities' idiosyncracies, such as city layout, may lead to biases in street view data for 28 cities across the globe, even when they are densely covered. We quantitatively uncover biases in the distribution of coverage of street view data and propose a method for evaluation of such distributions to get better insight in idiosyncracies in a cities' coverage. In addition, we perform a case study of Amsterdam with semi-structured interviews, showing how idiosyncracies of the collection process impact representation of cities and regions and allowing us to address biases at their source.

Figures (16)

• Figure 1: Overview of cities where the coverage distribution was evaluated. Cities were picked to ensure a good geographical spread.
• Figure 2: From left to right for Nairobi: Distribution of retrieved metadata in Google Street View, uniform coverage based on available drivable streets in OpenStreetMap, the difference of these two distributions $C_\Delta$ indicating the parts of the city that are oversampled or undersampled.
• Figure 3: Visualisation of how KL-Divergence and Earth Mover's Distance capture the differences between street view coverage distributions. Top left: For Singapore (GSV) the distribution is not uniform, but the over and undersampled areas are diffusely distributed throughout the city. Top right: For Dakar (Mapillary) coverage is almost only available in the western part of the city. The EMD is high because the centers of mass are distant. Bottom right: The coverage differences in Los Angeles (GSV) are contained to neighbourhoods, but the distrbutions have high overlap as the entire city has coverage. Bottom left: Reykjavik (GSV) has close to uniform coverage and the over and undersampled areas are diffusely distributed throughout the city.
• Figure 4: Coverage percentages plotted against the Earth Mover's Distance for Google Street View. Note that the Y and X axis are respectively extended and compressed for readability.
• Figure 5: Density plot ($C_\Delta$) of Google Street View Coverage in Los Angeles. Oversampled neighbourhoods such as Beverly Hills can have 12-14 images in a suburban street whereas undersampled neighbourhoods such as Compton may only have 4-5 images on similar streets.