Projects

Computer Vision for Ethnographic Research

Spring 2023
M.S. Urban Planning Thesis
Columbia GSAPP
Prof. Anthony Vanky

My master's thesis - can a machine read a storefront? A computer vision pipeline that scans street view imagery for the languages and colors of storefront signage, and tests whether they can flag neighborhood demographic change before official data sees it.

Scroll

The Question

Thesis
M.S. Urban Planning, 2023
Advisor: Anthony Vanky
Columbia GSAPP ×
University of Michigan
(OVPR Anti-Racism Grant)

3.56M street images
9 writing systems read
63,193 businesses

Sociology has long argued that when a neighborhood's people change, its businesses follow - Min Zhou's ethnic-entrepreneurship work built the canonical version of that claim. My thesis tests its converse: read the businesses instead - the languages on their signs, the colors they choose - and detect the people changing, earlier than any census table shows it.

The setting is the ethnoburb, Wei Li's name for the suburban ethnic business cluster. The instrument is computer vision over street view imagery, which quietly photographs every storefront in America every few years. If signage is legible at scale, the commercial landscape becomes a demographic sensor.

The Method, on a Detroit Street

The pilot ran in Madison Heights. Here the same reading runs across a Detroit corridor - Conant Avenue, in Banglatown

One block of Conant, read across five of street view's passes. A Bangla video store and a Korean taekwondo hall in the late 2000s give way to Yemeni service shops and Arabic storefronts by the 2020s - the kind of turnover the census records years late, if at all. Toggle the year, and drag to look around.

drag to look around

9723 Conant Ave, Hamtramck, read 2007 to 2025. Sign boxes are read at street level, colored by script. Imagery: Google Street View. The rest of this page is how the machine learned to do this.

Three Cities

The site research: where should a signage-reading method be born?

The thesis screened candidate cities on two axes: the vintage of their ethnic enclaves - New York's date to the early twentieth century, Detroit's largely to its second half - and their urban geography, from Manhattan's density to Los Angeles's polycentric sprawl to Detroit's low-rise suburban corridors. The proposal's original third city was San Francisco; Los Angeles replaced it by spring as the enclave literature and data pulled west. A method that reads storefronts has to work across all three landscapes.

Three trajectories, one method problem. New York holds near eight million; Los Angeles doubles; Detroit loses 65% of its people between 1950 and 2020, and its businesses moved to corridors like the one this thesis reads. City-proper decennial counts, U.S. Census.

Detroit

The Great Migration built the city's Black working class around the auto industry (Sugrue); Syrian, Lebanese, and Palestinian migration in the early twentieth century seeded what is now one of the largest Arab American populations in the country (Li). A Chinatown rose and fell; Japanese, Korean, and Vietnamese arrivals followed the industry after the war. Detroit's enclaves are the youngest of the three - late twentieth century, and suburban.

Enclave vintage: late 20th c. | Form: low-rise, corridor-bound

Los Angeles

The Latino share of the city grew from 10% in 1940 to 48% in 2019; the Asian share from under 1% to twelve points. The Chinese Exclusion Act of 1882 redirected migration into the Japanese, Filipino, and Korean communities that became Little Tokyo, Historic Filipinotown, and Koreatown (Wu). Freeway-era dispersal spread commerce into exactly the strip-mall form the pipeline reads (Scott).

Latino share 1940-2019: 10% → 48% | Form: polycentric

New York

The deepest enclave record: Chinatown dates to the mid-1800s, and the city's Asian population grew from 1.1 million in 2000 to 1.5 million in 2019, with Flushing and Sunset Park as its commercial anchors. Pyong Gap Min's comparison of Korean enclaves - Bergen County's suburban concentration against Queens's multiracial mix - is the clearest evidence that the form of an enclave shapes community power, not just its presence.

Asian population 2000-2019: 1.1M → 1.5M | Form: dense, vertical

Detroit went first: its enclaves are young enough that street view's archive covers their formation, its suburban strip-mall form puts signage at camera height, and the pipeline could be stress-tested inside a funded University of Michigan study of Asian and Arab American placemaking in the region.

The Search

3.5 million archived street photographs. Where do you start reading? Scroll.

One Corner, Five Years Apart

A Madison Heights strip mall as the pipeline sees it. Drag the handle.
2015 street view of a Madison Heights strip mall: a storefront flagged Vacant by the text detection pipeline ENCRICKET ENVACANT
2020 street view of the same strip mall: the vacant unit is now Ten Seconds Rice Noodle, with Bosna Grill beside it ZH过桥米线 · Ten Seconds Rice Noodle ENBOSNA GRILL

过桥米线 - Ten Seconds Rice Noodle - with Bosna Grill next door. July 2015 and November 2020, Metro Parkway, Madison Heights. Live panoramas: 2015 | 2020

Ten Seconds Rice Noodle storefront sign: Chinese characters above English text

The sign that changed the block: Chinese script set above English, the exact bilingual pattern the OCR pipeline is built to catch.

Live panoramas, not stills: the same spot in the pilot's two capture years. Drag to look around; the strip malls sit along the north side.

July 2015 and November 2020 · Metro Parkway, Madison Heights, MI.

The corridor holds more corners than this one. One Corner, Twelve Years follows a strip mall just west of here through every pass in the archive, 2008 to 2020.

The Pipeline

From every street in Detroit down to the text on one sign

Download

Every panorama arrives in pieces: a 26 x 13 grid of 512-pixel tiles, fetched tile by tile. The pilot pulled 10,002 panoramas this way.

One panorama = up to 338 tiles

Stitch

The tiles assemble into a single 13,312 x 6,656 equirectangular image - the camera's full sphere unwrapped flat at a 2:1 ratio.

88 megapixels per corner of the street

Crop

The top quarter is sky. The bottom quarter is asphalt. A low-rise corridor keeps its signs at eye level, so both are cut before any model runs - half the pixels never needed reading.

The decision that halves the compute

Read

EasyOCR walks the band in nine scripts, and detections lift off the storefronts with their confidences attached: HMART at 0.630, Chinese at 0.886, Korean at 0.671 - real values from the thesis output arrays.

Everything below 60% confidence is gated, but printed

Every stage below is a decision with a reason.

  1. 63,193 unique Detroit businesses in the Data Axle historic database via Wharton Research Data Services - 5.64M records over 25 annual snapshots, 1997-2021. Decision: ground-truth the imagery against business records, because signage reading means nothing without knowing what was actually there.
  2. 1,319,464 street coordinates generated by interpolating every 10-20 meters between OSMnx street-network nodes. Decision: Google's API only serves the latest panorama per location, so a custom scraper against the undocumented photometa endpoint recovered the historical archive back to 2007 - 3,561,914 panoramas inventoried, 10.3% of metadata calls returning nothing, and the retry-wrapped scraper closing the metro sweep with an empty error log.
  3. 9,237 → 577 pilot funnel: panorama IDs inside the Madison Heights test bounds, filtered to 8,734 valid frames, 633 after manual cleaning, 577 in high resolution. Decision: explicit exclusion rules - vehicles with text decals, residential frontage, advertising not attached to an establishment - and the top and bottom 25% of each frame cropped away, because the corridor is low-rise and sky and asphalt carry no signs.
  4. 4 OCR engines tested across the project's arc: the proposal specified EAST + Tesseract; an eight-language Tesseract pass over 602 stitched panoramas (58,720 detections) showed Bengali costing ~245 seconds per panorama against English's ~32 and forced the switch; the thesis shipped EasyOCR across nine scripts with an adapted pretrained ABINet (Fang et al., CVPR 2021) and stroke-width-transform code compiled for comparison. Decision: engines are disposable, the comparison isn't.
  5. 60% confidence gate on every detection, with an ASCII check splitting Latin from Asian and Middle Eastern script pipelines and langdetect cross-verifying the language call. Decision: publish the low-confidence tables too - the thesis prints its 0.02-confidence failures next to its successes.

Groundwork: CITI human-subjects certification, October 2022; proposal defended December 2022; thesis jury April 12, 2023.

What the Pilot Read

Language and color signals from 577 images of two strip malls
Arabic3.7%
Chinese2.0%
Korean1.3%
Share of confidently detected text by language across the pilot imagery. Small numbers, real signal: the corridor's Arab and Asian commercial presence shows up in machine-read signage, in proportions that track the area's business mix.
H Mart storefront panorama with cropped detections of its trilingual signage: English, Chinese, and Korean
One H Mart, three scripts: the same grocery announces itself in English, Chinese (超市), and Korean (마트). Each crop is a real detection from the pipeline.
Color Thief reads the H Mart sign: dominant deep red, a color carrying prosperity connotations across East Asian commercial design.Swatches are the pipeline's output values. Color proved noisier than language, and the thesis says so.

Findings, Limits, and What Comes Next

What the pilot proved, where it stops, and the tests still open
  • Street view is a usable ethnographic instrument. The pipeline extracts storefront language and color from ordinary street imagery at scale, across nine scripts - the core feasibility claim, demonstrated.
  • Language is the strong signal. Script shares in machine-read signage tracked the pilot corridor's actual business mix. Color was noisier: backgrounds and shadows skew extracted palettes, and the thesis documents that rather than hiding it.
  • It only works alongside other data. Signage reading needs ground truth - historic business records, census tables - to mean anything. It is a leading indicator, not a standalone measure.
  • The archive is the constraint. Historical street imagery thins out fast before the mid-2010s. Tracking substantial change needs more years of coverage than most corridors have - a limit that relaxes every year the archive grows.
  • Some change is invisible to the camera. Self-orientalization - businesses performing an ethnic identity for the market - and gentrification that rebrands in the same script read identically to organic change. The method sees signs, not intent.

Where it stops

  1. The archive has a floor Street view thins out fast before the mid-2010s. Long-horizon change is only measurable where Google drove early; the pilot corridor's record starts in 2007, and most residential blocks start later.
  2. The ground truth has gaps Data Axle misses the smallest and youngest businesses, and listings lag reality. Records were manually verified and labeled before they were allowed to ground anything.
  3. The archive is sometimes silent 10.3% of metadata queries returned nothing at all. A coverage map has to mark where the record is empty, not just where it speaks.
  4. 577 images is a pilot, not a study Two strip malls, one corridor, one time pair. The pilot demonstrates that the reading works; it does not describe Detroit, and the page before this one is careful to never claim it does.
  5. Color is noisier than language Extracted palettes absorb brick, shadow, and sky along with the sign. Language carried the pilot's signal; color needs masking work before it can carry anything.
  6. The failures are printed Detections scoring 0.02 sit in the thesis tables next to ones scoring 0.99. The 60% confidence gate is a documented choice, not a hidden one.

The tests still open

  • The prediction test. The pilot validated the reading - languages on storefronts, extracted at scale. The experiment it sets up but does not run is the correlation: do signage shifts lead demographic change in the official record? That test is the method's natural next step, and it is still open.
  • Two more cities. The three-city design above is the map for it: the same pipeline pointed at Koreatown or Flushing, where enclave vintage and urban form differ from Detroit's corridors in exactly the ways the site research laid out.
  • The workaround became a product. In 2023 the historical archive was only reachable by scraping an undocumented endpoint. Google has since shipped official historical Street View access and a Street View Insights API aimed at exactly this kind of temporal analysis - the hardest engineering in the pilot is now a product surface, which lowers the cost of every test above.

Reading a neighborhood is not a neutral act

Pointing computer vision at immigrant neighborhoods carries the risks the thesis names directly: the same pipeline that describes commercial change could feed surveillance or predictive-policing systems, and a model's reasoning is hard to put in front of a community meeting. That is why the tool is framed for planners as descriptive, never predictive-policing adjacent.

This corner spoke English in 2015 and answered in Chinese by 2020, and no table anywhere records the change. Only the pictures do. That is what street view is, underneath: an accidental archive of the storefronts data forgets. Reading it carefully is a way to count them before they are gone.

The Machine, End to End

Every cut the pilot made, stacked
From 3.56 million catalogued panoramas down to the 577 frames the OCR actually read - each ledge is one of the methodological cuts described above. What came out the top: Arabic at 3.7% of confident text, Chinese at 2.0%, Korean at 1.3%, in a corridor whose businesses say exactly that.
Master's thesis for the M.S. Urban Planning program at Columbia GSAPP, May 2023, advised by Prof. Anthony Vanky. Supported by Columbia GSAPP and a University of Michigan OVPR Anti-Racism Grant ("Asians in Southeast Michigan: Information Networks and Placemaking as Responses to Racist Violence"). Street imagery via the Google Street View APIs; historic business records from Data Axle via Wharton Research Data Services.

Kirthi Balakrishnan

Advisor: Prof. Anthony Vanky, Columbia GSAPP