This project employs machine learning and image data to build a model for predictive image prefernece. Images are sourced from the google street view api and classifies them into buckets one, on left(or top) and zero on right(or bottom). In this application of the build, 10 images from each bucket are randomly placed next to each other. These images were classified by an algorithm not a person.

I labeled a training set of images about 150 for each bucket as one, or zero. Images placed in bucket one generally had high contrast and a facade of a house was dominate. Images placed in bucket zero are low contrast or obscured by rain on lens. The image is dominated by the street, cars, brush or other imagery that was not a facade of a structure. Transfer learning was used with the Inception v3 Convolutional Neural Network with tensorflow to create a model locally. The generated model can classify images not in the training set as one or zero. 754 images were classified and the data is stored in a JSON file. Every 10 days images are pulled to the page via the google street view api. I am using a version of this tool to select images from Oakland, CA to create a body of paintings. Oakland Street View.

New machine learning technology toolkits like tensorflow combined with vast amounts of data(images from google street view) gives rise to complex reality thresholds. Images can be transformed into meta data that does not have to be seen visually to be categorized. The preferences of the user is captured in the algorithm.

References
Fei Fei Li TED talk 2015
Tensorflow Retrain image classifier
Image net
Google Street View API