Property Images Are Worth More Than 1000 Words


Image data is one of the most common types of data captured and analyzed for insights in organizations today. Modern Machine Learning techniques have simplified image analysis to the point where its been adopted in nearly every domain. In one of our recent projects with Bradford Technologies, a leading real estate appraisal software provider, we explored methods to build intelligent appraisals models through the use of image-based, deep learning technology.

The Problem

The world of real estate appraisals has become incredibly data driven. Web-based, property appraisal companies such as Zillow or Redfin are becoming more popular. However, combining machine expertise with human appraiser expertise results in more accurate appraisals.

While appraisal reports already process many metrics to determine the estimated appraisal value, there are always more features that can be used as input to improve the appraisal model’s accuracy. Bradford was specifically interested in generating new features using their database of property images to further improve their tool suite.


In order to get started, we needed to select an initial feature to engineer based on the available image data. Features such as roof type/quality and lawn type were discussed but we eventually settled on garage doors. Garage doors typically have a consistent shape across properties making them a good first feature.

POC using Garage Doors

To jump start our project, we leveraged transfer learning with the Tensor Flow Inception model as our base. In this way, we did not have to build and train a whole deep learning model, just the last layers of one. We started with just single-car, garage doors and then augmented the model to also detect two-car, garage doors. After hyper-parameter tweaking and determining we needed a confidence threshold of 85%, we were able to produce some great results. Our model specifically had an accuracy of 89.4% and a precision of 87.6%.

Improving Results

Most of the images we used to train and validate our models were small in size (640x480). We hypothesized that the small images provided less signal for the model to pick up on and that if we use larger images we’d have better results. To quickly test this hypothesis, we ran five larger images and saw the model’s results improve (see below). Assuming a dataset with larger images becomes available, we’ll rebuild the model and further evaluate these initial results.


Using Bradford’s existing property image database, we successfully extracted the garage door type and count features for each property. Additionally, because the model returns the garage door location in the image, the relative position of the garage doors could be a useful model feature. This demonstrates that additional features can be engineered from property images to be used in appraisal models.

If you’d like to learn more about this project, feel free to reach out to us at