Abstract
In this work, we build a series of machine learning models to predict theprice of a product given its image, and visualize the features that result inhigher or lower price predictions. We collect two novel datasets of productimages and their MSRP prices for this purpose: a bicycle dataset and a cardataset. We set baselines for price regression using linear regression onhistogram of oriented gradients (HOG) and convolutional neural network (CNN)features, and a baseline for price segment classification using a multiclassSVM. For our main models, we train several deep CNNs using both transferlearning and our own architectures, for both regression and classification. Weachieve strong results on both datasets, with deep CNNs significantlyoutperforming other models in a variety of metrics. Finally, we use severalrecently-developed methods to visualize the image features that result inhigher or lower prices.