Computational social science · street-level imagery · India

Can a street photo measure a city’s infrastructure — and can we trust which streets got photographed?

UrbanLens India asks two questions at once, deliberately. Measuring sidewalks, paved roads, and street lighting from crowdsourced street-level imagery is only useful if we also know how unevenly that imagery covers the country it claims to describe.

Read the methods Browse the city pool

Question 1

Can street-level computer vision measure real infrastructure?

Sidewalk presence, road surface, street vegetation, lighting, and building frontage density — five attributes with direct precedent in the urban-informatics literature — scored from single street-level frames sourced from Mapillary. See /vision for the target attributes and /generalization for how it’s tested across cities the model has never seen.

Question 2

How biased is that measurement, geographically?

Street imagery coverage is not random: it follows where contributors already drive with a camera. Cities were selected before checking imagery availability, so coverage can be measured as an outcome and corrected for, not used to cherry-pick well-photographed places. See /bias for the missing-data framing.

As of now

What’s actually measured today

Candidate cities in the study pool

docs/city_selection.md

Cities with a built OSM road network

data/interim/road_universe_meta.json

1,034,197

Road-network sample points (50m spacing)

data/interim/road_universe_meta.json

Target infrastructure attributes

data/labels/schema.json

Not shown above because it does not exist yet: Mapillary imagery-coverage percentages, trained-model accuracy, and infrastructure estimates. Those sections say so plainly — see /coverage and /infrastructure.

Explore

CitiesThe 14-city candidate pool, stratified by region, population, and economic tier.View →CoverageHow imagery coverage is measured, and why it isn't a selection filter.View →InfrastructureNaive vs. coverage-reweighted infrastructure estimates.View →Bias LabMCAR/MAR/systematic missing-data framing, propensity modeling, masking experiments.View →VisionThe CV benchmark: target attributes, candidate models, metrics.View →GeneralizationRandom split vs. spatial split vs. city holdout.View →MethodsThe full methods write-up.View →DataSources, licenses, and retrieval dates for every dataset used.View →PaperThe write-up, once it exists.View →