Cathedral of Cold

Feb 20, 2026 · idea

A map of where the city exhales refrigerated sadness.

The MVP is simple: find outdoor AC units / vents, guess where their cold breath goes, then draw it on a map like it’s a weather forecast for bad decisions.

What it does (MVP)

• Detect outdoor AC units and vents in street-level/facade images.
• Estimate facade orientation from building footprints (approx. “which way this wall faces”).
• Combine with wind direction to draw a plausible “plume direction” vector.
• Produce a map layer of:
  • detections (points)
  • “cold breath” vectors (lines)
  • optional density overlay (a heatmap of comfort addiction)

Why it exists

Because infrastructure is real, invisible, and strangely poetic when you chart it.

Also because nothing says “modern city” like spending energy to fight physics and then venting it into the street.

Plan and steps (so this isn’t just a moodboard)

Phase 0 — Feasibility (cheap reality check)

Goal: prove we can detect anything at all before we start buying GPU time like it’s candy.

• Define classes: condenser unit, vent/exhaust grille, split AC outdoor unit, etc.
• Baseline run with an off-the-shelf detector (or quick fine-tune if needed).
• Decide scope: one street / one district / one city.

Deliverable: a small demo map with 50–200 detections and “good enough” placement.

Phase 1 — Data acquisition (the part everyone pretends is easy)

This project needs two kinds of data:

1) GIS layers (usually manageable)

• Building footprints (and height if available)
• Addresses/parcels (optional)
• Any public layers you can legally use (ArcGIS portals are often the source of truth)

2) Facade imagery (the actual hard part)

Outdoor units are typically not visible from above, so we need street-level / facade imagery.

Options:

• Open/crowdsourced imagery datasets (license constraints matter)
• Existing street imagery if permitted
• Your own capture (phone + GPS) if you want clean licensing and full control

Deliverable: an image set that can be legally used for detection + mapping.

Phase 2 — Labeling (aka “how much do you like clicking boxes?”)

• Write a labeling guide: what counts as “AC unit” vs “random electrical box”
• Label a starter set (ballpark: hundreds to a few thousand instances)
• Split train/val/test to avoid training a model that only works on one street forever

Deliverable: a dataset that can actually support a fine-tune.

Phase 3 — Training (where the GPU money goes)

• Fine-tune an object detector
• Validate on new streets/areas (generalization check)
• Iterate until the model stops hallucinating vents in potted plants

Deliverable: a model that detects units at usable precision/recall.

Phase 4 — City-scale inference + mapping

• Run inference across the image corpus
• Geolocate detections (depends on imagery source)
• Estimate facade orientation from building footprints (edge normals as proxies)
• Pull wind direction for a chosen time window (or use average wind roses)

Deliverable: map layers and a first interactive map.

Phase 5 — Publish + ship

• Publish an interactive web map embed
• Write-up with transparent methodology + known failure modes
• Optional: downloadable tiles / layers if licensing allows

Budget (why €3,000 is not just vibes)

Funding target: €3,000.

It’s intentionally small enough to be achievable, and big enough to cover the parts that are boring but real:

1) Compute: ~€1,200

Training + reruns + large inference passes.
GPU pricing varies wildly, so think “a meaningful chunk of hours” rather than one magic number.

2) Labeling: ~€1,200

Either you pay for annotation, or you pay in time.
This budget line is basically: “do we want this in weeks, or in months?”

3) Data / hosting / misc: ~€600

Storage, map tiles, small tooling costs, and the inevitable “why does this format exist?” tax.

If we collect imagery ourselves, some of that shifts to travel/time instead of cash. The point is transparency: the money buys progress, not mystery.

Known risks (a.k.a. the truth)

• Imagery licensing can kill the “open dataset” dream fast. We’ll keep it legal.
• Geolocation depends on image source. If it’s messy, mapping gets fuzzy.
• Facade orientation is an approximation. It’ll be “plausible,” not “physics-grade.”
• The map will include false positives. That’s part of the aesthetic and also why we validate.

Sponsoring this

If you sponsor, you get:

• attribution on this project page
• (optionally) influence via constraints (“do it for this district”, “use only open data”, “no manual labeling”, etc.)

What you don’t get:

• editorial control
• paywalled output
• a guarantee that the city will stop being ridiculous

The work stays public. The sadness stays refrigerated.