DualSight: trained models

Trained models for DualSight, which estimates the price of a home in Austin, Texas from a Sentinel-2 satellite image of the neighbourhood together with the listing details.

Files

dualsight_trained_model.zip contains:

Model	What it is
Neural hybrid	ResNet50 image branch (ImageNet weights, last layers fine-tuned) + dense numeric branch, trained end to end on log(price)
Deep + boosting hybrid	ResNet50 image features compressed with PCA (32 components) + the listing numbers, in a gradient-boosting model

Plus the image features, reference predictions used to verify the import, held-out test predictions and training history.

Numeric inputs: living area, year built, bedrooms, bathrooms, average school rating, lot size.

Results

Scored on the same 2,994 held-out Austin homes, none used in training:

Model	Typical error	Within 20% of listing
Deep + boosting hybrid	12.4%	69.9%
Neural hybrid	12.7%	68.6%
Numbers only, no image (for comparison)	15.7%	59.8%

How to use

Follow the setup in the DualSight repository, download the zip, then:

python pipeline/import_kaggle_model.py path/to/dualsight_trained_model.zip

The import rebuilds the neural hybrid in your local TensorFlow, checks its predictions against the reference ones, and retrains the gradient-boosting step with your own scikit-learn.

Limitations

• Trained on Austin, Texas listings only; it will not transfer to other cities without retraining.
• Least accurate for the cheapest and the most expensive homes (estimates are pulled toward the middle of the range).
• Listing prices are not sale prices.