Let's see if I can make a general purpose aerial imagery vector, then use it to predict how much I want to live in a place.

1. Python Setup
2. Getting Raw Data
3. Generating Our Data
4. Running the Analysis

This project uses Python 3.12, but it would probably work with other versions of Python 3.

python3.12 -m venv ./venv/ && \
source ./venv/bin/activate && \
pip install -r requirements.txt

This analysis uses the Inria Aerial Image Labeling Dataset, available at the link on 22 Apr 2024.

1. Download the multipart 7z files by running (or referencing) scripts/download-data.sh
   • Alternative: curl -k https://files.inria.fr/aerialimagelabeling/getAerial.sh | bash
   • But you shouldn't pipe scripts from the internet to your shell.
   • Make sure you're in the ./data/ directory.
2. Run or reference scripts/extract-data.sh to extract the archive
3. Carry on

We're going to make little sub-tiles from our images, and associate them with other information (including parent image and image group and a score).

python scripts/generate-subimages.py

You've got your virtual environment, installed the deps, made the data.

export PYTHONPATH="${PYTHONPATH}:$(pwd)/src"
python src/pca.py && \
python src/train.py && \
python src/eval.py

If your results look like mine, they aren't useful.

Why didn't this work well?

• Maybe our input data vs training data isn't the same zoom level, color space, etc...
• I'm only extracting color distribution, not the things that actually characterize a place
• Maybe you can't know enough "ground truth" about a place by the way it's built from overhead

If I wanted to try again - better but slower - I could extract features from images like curviness of roads, angles on buildings, rooftop coloration, road size vs building size, etc...