[MS/PhD] Paid Student Worker: Computer Vision on School Facilities (USC Economics)

The following announcement is from USC Dornsife, Department of Economics- Prof. David Schönholzer (dschonho@usc.edu). Please contact them directly if you have any questions.

The project. American school districts spent $90 billion on school construction and renovation in 2019-20, about $1,760 per student. What that money physically buys is only partly recorded: enrollment, grades, and location are in federal data, but the buildings themselves are not. There is no national record of which schools have air conditioning, how many classrooms they hold, whether the gym was rebuilt in 2015, or which campuses are fenced. We want to build that record for all ~130,000 U.S. public schools by combining street view imagery, satellite and aerial imagery, open building-footprint data (Microsoft’s US Building Footprints, Overture Maps), and administrative data.

Concretely, we want per-school measures of:

• Building footprint and campus area; estimated classroom count and seat capacity
• HVAC signatures (rooftop units, chillers, window units)
• Rooftop solar
• Athletic facilities: tracks, fields, courts, pools
• Auditoriums, gymnasiums, and other large-span structures
• Security features: perimeter fencing, gates, controlled access points
• Portable and modular classrooms
• Visible condition and evidence of recent renovation

Where possible we want these measures to be time-varying rather than a single snapshot. Historical Street View panoramas reach back to roughly 2007 in many locations, and repeat satellite coverage is denser still, so a well-built pipeline should be able to date when a solar array, a portable classroom, or a new wing first appeared.

These map onto the capital-spending categories in my Quarterly Journal of Economics paper on school bonds (classrooms, infrastructure, HVAC, safety and health, STEM, athletics, land, transportation), where we could measure what districts spent money on but not what physically changed. This project closes that gap.

The role. 15 hours per week at $20/hour, USC student worker appointment, starting mid-September 2026 and running through the fall semester, with renewal likely if it goes well. Hybrid; some in-person meetings on UPC. You would own the imagery pipeline end to end: acquisition, model selection and fine-tuning, validation against ground truth, and cost control at scale. You would meet with me weekly.

Who should apply. MS or PhD students in Computer Science or a related program with hands-on experience in computer vision (detection, segmentation, or VLM-based extraction) and comfort building data pipelines in Python. Experience with geospatial data, remote sensing imagery, or Street View APIs is a plus but not required. What matters most is that you can get a measurement pipeline working and deliver high-quality, transparent work at a fast pace.

How to apply: two stages.

Stage 1, due Friday, August 21. Email dschonho@usc.edu with subject line “School Facilities CV – [Your Name]” and include:

1. Your résumé.
2. A link to one thing you have built (repo, demo, or paper) plus two sentences on what part was yours.
3. Under 200 words on which of the measures above you think is hardest to extract from imagery, and why.

Stage 2, by invitation. Shortlisted applicants receive a short take-home: a list of 25 real schools with coordinates, and a request to produce a structured measurement of four attributes for each, with a per-field confidence, plus a one-page memo on how you validated it and what your error rate is. If you have time, show one attribute measured at two different dates. It is designed to take about five hours, and free API tiers are sufficient at this scale. Please do not spend money on it. I will read your validation section more carefully than your accuracy number.

Decisions by September 11.

Questions welcome at dschonho@usc.edu.

Published on August 3rd, 2026Last updated on August 3rd, 2026