Streetscope: Measuring Gender Representation in Public Spaces

Analyzing classified person-sightings in GoPro wearable-camera imagery collected along routes in four Indian cities.

Key Findings

The tables below are regenerated by scripts/10_analysis.py (via make figures/make tables).

Summary

City	Images	Pedestrians	Female share [95% CI]	Sex Ratio (F/1000 M)
Mumbai	2,809	12,331	19.1% [17.4%, 20.8%]	236
Navi Mumbai	2,189	5,186	18.1% [16.0%, 20.2%]	221
Bangalore	2,015	2,311	27.4% [25.7%, 29.1%]	377
Delhi	2,417	3,020	19.8% [17.3%, 22.2%]	246

By Mode

Mode	Mumbai	Navi Mumbai	Bangalore	Delhi
Pedestrians	19.1%	18.1%	27.4%	19.8%
Two-wheelers	8.2%	5.3%	8.8%	5.9%

By Road Type

City	Primary	Secondary	Tertiary	Residential
Mumbai	20.3%	16.9%	18.6%	20.3%
Navi Mumbai	16.7%	15.9%	18.1%	23.3%
Bangalore	28.8%	22.8%	25.1%	29.4%
Delhi	16.5%	16.8%	19.6%	23.4%

Women-coded pedestrian sightings are below parity in the collected imagery in every city, and the female share is lower among two-wheeler riders than among pedestrians. Census sex ratios are population quantities and are not directly comparable to this route-image sample.

Pipeline

The project is a linear pipeline from sampling design to published results:

geo-sampling → itineraries (allocator) → GoPro collection → EXIF/GPS/frame extraction
  → face-frame extraction → human annotation (Label Studio)
  → parse annotations → assign GPS → enrich with road type → analysis_data
  → figures + tables + maps

Frozen vs. fair-game

The collection half is frozen: it has been run once in the field and on the raw footage, and its outputs (extracted frames, EXIF, GPS index, annotations) are treated as immutable inputs. The analysis half is fair game: it transforms those frozen inputs into the published dataset and outputs and can be re-run freely.

Stage	Script(s)	Status
Geo-sampling / itineraries (allocator)	scripts/gen_data/<city>/run_*_sampler.py	frozen
Video processing: EXIF, GPS, frame extraction	00_process_videos.py	frozen
Face-detection frame extraction	01_extract_face_frames.py	frozen
Coverage QA / frame compression	02_plot_video_coverage.py, 03_compress_frames.py	frozen
GPS lookup index / EXIF recovery	04_build_gps_index.py, rebuild_csvs_from_exif.py	frozen
Human annotation (Label Studio)	external	frozen
Parse annotations	05_parse_annotations.py	fair game
Assign GPS to frames (ts/lat/lon)	06_assign_frame_gps.py	fair game
Enrich with road type (itinerary + OSM)	07_enrich_with_geo.py	fair game
Build analysis dataset	08_build_analysis_data.py	fair game
EDA / publication figures+tables / maps	09_eda.py, 10_analysis.py, 11_make_maps.py	fair game
Descriptive clustered inference	inference.py	fair game
LLM annotation comparison (optional diagnostic)	12_validate_annotations.py	fair game
Inter-rater reliability	13_interrater_reliability.py	fair game
Descriptive pattern mining	14_descriptive_patterns.py	fair game

Run the analysis half

The analysis half starts from the frozen GPS index and annotations and never re-touches videos or frames:

# One city, end to end (analysis_data + its figures/tables/maps), using cached GPS index:
make analyze CITY=mumbai

# Regenerate figures/tables/maps across cities from existing analysis_data:
make figures CITIES=mumbai,navi_mumbai

# Tables only:
make tables CITIES=mumbai,navi_mumbai

Equivalent direct invocation:

.venv/bin/python scripts/run_pipeline.py --city mumbai --skip-rebuild-gps

Re-running collection (frozen — only if raw footage changes)

# Re-process videos and re-extract frames (requires video_dir in cities.yaml):
.venv/bin/python scripts/run_pipeline.py --city mumbai --process-videos

Inference

The primary estimand is the share of classified pedestrian sightings coded as women: sum(women pedestrians) / sum(all pedestrians). This definition is also used in the Google Street View analysis. Combined pedestrian/two-wheeler estimates and image-level means are secondary.

The extracted images are face-triggered frames along collected routes, not a probability sample of streets or residents. Results therefore describe visible classified sightings in the collected imagery. Confidence intervals use a collection-day cluster sandwich with a t distribution on G−1 degrees of freedom; they quantify variability across the observed fieldwork days and do not imply population representativeness.

Counts recorded as 10+ are interval-censored. The primary tables use 11, the known minimum, and report sensitivity to replacement values 15, 20, and 30.

The optional OpenAI comparison is a diagnostic, not a gold-standard validation. Install it with uv sync --all-extras; the script defaults to the pinned image-capable model snapshot gpt-4o-2024-11-20 and performs no request unless --submit or --check is supplied.

Setup

# System tools (macOS)
brew install ffmpeg exiftool

# Python environment (uv-managed .venv):
make install

make install installs the pinned analysis and development dependencies from uv.lock.

Optional: OSM ground-truth road type

07_enrich_with_geo.py can attach an OSM "ground truth" road type alongside the itinerary road type. It is skipped automatically unless the city's PBF is present. Download the region from Geofabrik into data/osm/ using the filename in cities.yaml, e.g.:

data/osm/maharashtra-latest.osm.pbf   # mumbai, navi_mumbai
data/osm/karnataka-latest.osm.pbf     # bangalore

When present, tableS1_road_type.tex gains an OSM ground-truth section.

Directory Structure

scripts/
  gen_data/{city}/           Sampling-design scripts (frozen)
  00_process_videos.py ...   Pipeline scripts (see table above)
  inference.py               descriptive collection-day clustered inference
  run_pipeline.py            Analysis-half orchestrator
data/
  {city}/
    exif/                    Raw EXIF text files (frozen)
    gps_index/               GPS lookup parquet files (frozen)
    labelstudio/             Label Studio JSON exports (frozen)
    face_frame_metadata.csv  Face-detection frame log
    analysis_data.parquet    Final analysis dataset (primary annotator)
    analysis_data_long.parquet  Long format (all annotators)
  annotation_task/           Extracted frames for annotation
  osm/                       Optional OSM PBFs for ground-truth road type
sampling/
  {city}/itineraries/        Itinerary road-type classifications
figs/                        Generated figures and maps
tabs/                        Generated LaTeX tables
cities.yaml                  Per-city config (video_dir, osm_file, timezone)
Makefile                     Analysis-half workflow targets

Outputs

Data

• data/{city}/analysis_data.parquet - Primary analysis dataset
• data/{city}/analysis_data_long.parquet - Long format (all annotators)

Figures (figs/)

• fig2_distribution.pdf - Proportion female distribution
• fig3_multipanel.pdf - Multi-panel summary (mode, road type, POI, time)
• fig4_weekday_weekend.pdf - Weekday vs weekend comparison
• fig5_pedestrian_crowdsize.pdf - Female share by pedestrian crowd size
• map_locations_{city}.pdf / .html - Data collection locations
• map_sexratio_{city}.pdf / .html - Female share by ~500m cell
• map_overview_india.pdf - All cities on one India basemap
• eda_*.pdf - Exploratory analysis plots

Tables (tabs/)

• table1_city_summary.tex - City-level summary
• tableS1_road_type.tex - By road type (itinerary + optional OSM ground truth)
• tableS2_temporal.tex - Temporal patterns
• tableS3_poi_infrastructure.tex - POI and infrastructure
• tableS4_irr.tex - Inter-rater reliability (via make irr)
• descriptive_patterns.md - Accompaniment, place rankings, joint regression

Data Coverage

Metric	Mumbai	Navi Mumbai
GPS coverage	97.4%	99.4%
Itinerary match	66.3%	72.1%
Hour range (IST)	7-19	9-18

Bangalore and Delhi are annotated and included in the analysis outputs. Hyderabad has sampling itineraries and partially processed footage (face frames extracted); annotation/analysis for it is in progress.

Per-frame counts are top-coded at 10 ("10+" recorded as 11), and collection spans roughly 06:00-22:00 IST only.

License

MIT