Image Generation & Annotation

Moderate · Data & Analysis track · ~35 min hands-on + readings and quiz

← Back to the Data & Analysis track

Make communication graphics responsibly, and turn images into labeled data.

What you’ll be able to do

  • Generate communication graphics and critique them for accuracy and representation
  • Annotate images (bounding boxes, COCO) for analysis
  • Disclose AI-generated visuals and apply consistent labels

Overview

Generative tools make graphics fast; annotation tools turn images into labeled data you can measure (street features, hazards, conditions). Both need judgment — generated images can mislead or misrepresent, and labels need consistent definitions.

Disclose AI-generated visuals in published materials, and check that people and places are represented fairly.

Practice activities

Activity 1 · Moderate — Generate and critique a safety graphic

Time ~17 min · Tools Adobe Firefly / Express (Drexel Moderate-Risk data)

Goal. Make a communication graphic responsibly.

Setup. No dataset needed. Open Firefly or Adobe Express.

Steps.

  1. Generate an image:

    Create a clean, inclusive illustration for a pedestrian-safety campaign: people of varied ages and abilities crossing safely at a marked crosswalk in a city. Flat, friendly style, with space for a headline.

  2. Critique what comes back:

    • Is anything inaccurate or unsafe (e.g., people crossing against the signal)?
    • Who is represented, and who isn’t?
    • What would you disclose when publishing it?

  3. Regenerate to fix one issue.

Expected result. A usable draft graphic plus notes on accuracy, representation, and disclosure.

Check your work. Ask yourself: could this image mislead anyone about safe behavior? If so, fix it before using it.

Common pitfalls. Generated scenes can show unsafe or unrealistic situations — review carefully. Always disclose AI-generated visuals in published materials.

Stretch (optional). Make a second version adapted for a different neighborhood or audience.

Activity 2 · Moderate — Annotate street images

Time ~20 min · Tools Label Studio (free) + a browser

Goal. Turn images into labeled data you could measure.

Setup. Use these 10 publicly available Philadelphia intersection images from Google Street View (Google automatically blurs faces and license plates). Open each link and save the image, or right-click to copy:

  1. Broad & Lehigh — wide arterial, no visible crosswalk markings
  2. Broad & Walnut — downtown, marked crosswalks and curb ramps
  3. Frankford & Allegheny — complex intersection
  4. 5th & Spring Garden — near highway ramp
  5. Lancaster & 40th — West Philadelphia
  6. Oregon & Broad — South Philadelphia
  7. Girard & Front — Fishtown
  8. Roosevelt Blvd & Adams Ave — high-crash corridor
  9. Market & 52nd — West Philadelphia commercial
  10. Cottman & Bustleton — Northeast Philadelphia

If the Street View links don’t load at the exact angle, navigate to the intersection in Google Maps and drop the Street View pegman. Install or open Label Studio (free, open-source).

Steps.

  1. Set up a bounding-box project with two labels: marked crosswalk and curb ramp. Write a one-line definition for each before you start.

  2. Label all ~10 images.

  3. Export to COCO format.

  4. Sanity-check consistency with AI:

    Here is my labeling guide and a sample of my annotations. Point out where my label definitions are ambiguous or applied inconsistently.

Expected result. Ten annotated images exported as COCO, plus a tightened labeling guide.

Check your work. Re-label 2 images without looking at your first pass. If you disagree with yourself, the definitions need work.

Common pitfalls. Vague definitions wreck any downstream analysis. Never upload images with identifiable people to an unapproved tool.

Stretch (optional). Add a third label (e.g., “missing curb ramp”) and note how it changes how consistently you can annotate.

Check your readiness

Answer these, then check — your score suggests whether to dive in or skim the readings first.

Useful resources

  • Adobe Firefly — commercially-safe image generation (Drexel Moderate-Risk data).
  • Label Studio — open-source image and text annotation.
  • CVAT — computer-vision annotation tool.