[Paper Review] The Price is Right: Predicting Prices with Product Images.

Steven Chen, Edward Chou+1 more|arXiv (Cornell University)|Mar 29, 2018

Market Dynamics and VolatilityEconomics, Econometrics and Finance12 references4 citations

TL;DR

This paper proposes a deep learning approach to predict product prices from images using custom datasets of bicycles and cars. It employs ensemble CNN models with transfer learning and visualizes salient image features influencing price predictions, achieving superior performance over traditional methods like HOG and SVM baselines.

ABSTRACT

In this work, we build an ensemble of machine learning models to predict the price of a product given its image, and visualize the features that result in higher or lower price predictions. We collect two novel datasets of product images and their MSRP prices for this purpose: a bicycle dataset and a car dataset. We set baselines for price regression using linear regression on histogram of oriented gradients (HOG) and convolutional neural network (CNN) features, and a baseline for price segment classification using a multiclass SVM. For our main models, we train several deep CNNs using both transfer learning and our own architectures, for both regression and classification. We achieve strong results on both datasets, with deep CNNs significantly outperforming other models in a variety of metrics. Finally, we use several recently-developed methods to visualize the image features that result in higher or lower prices.

Motivation & Objective

• To develop a machine learning system that predicts product prices from visual features in images.
• To collect and release two novel datasets of product images with their MSRP prices for research.
• To compare deep learning models against traditional methods like HOG and SVM for price prediction.
• To interpret model decisions by visualizing image regions that contribute most to high or low price predictions.
• To evaluate performance across regression and classification tasks using multiple metrics.

Proposed method

• Collects two new datasets: one of bicycles and one of cars, each with images and corresponding MSRP prices.
• Uses histogram of oriented gradients (HOG) and pre-trained convolutional neural network (CNN) features as baseline inputs for regression and classification.
• Trains multiple deep CNNs using both transfer learning and custom architectures for both price regression and price segment classification.
• Employs ensemble modeling to combine predictions from multiple CNN architectures for improved robustness and accuracy.
• Applies explainability techniques to visualize image regions that positively or negatively influence price predictions.
• Evaluates models using standard regression and classification metrics on both datasets.

Experimental results

Research questions

• RQ1 Can deep CNNs effectively predict product prices from images alone, outperforming traditional handcrafted features like HOG?
• RQ2 How do different CNN architectures and training strategies impact price prediction accuracy on real-world product images?
• RQ3 Which visual features in product images most strongly correlate with higher or lower price predictions?
• RQ4 To what extent do transfer learning and custom CNN designs improve performance on novel product price prediction tasks?
• RQ5 How interpretable are the predictions made by deep learning models in the context of pricing?

Key findings

• Deep CNNs significantly outperform baseline models such as linear regression on HOG features and multiclass SVMs in both regression and classification tasks.
• Transfer learning with pre-trained CNNs yields strong performance, demonstrating the value of pre-existing visual representations for price prediction.
• Custom CNN architectures trained from scratch also achieve high performance, indicating that domain-specific learning is effective.
• Visualizations reveal that features such as brand logos, vehicle make/model, and product condition are key drivers of price predictions.
• The ensemble of deep learning models achieves state-of-the-art performance on both the bicycle and car datasets in terms of regression and classification accuracy.
• The interpretability methods successfully highlight image regions—such as wheels, headlights, or frame details—that correlate with price variations.