The goal of this project is to develop Deep Learning models that can predict the price of a product, such as a TV set or a car, based on an image acquired by a customer.
One of the challenges when handling data coming from two different sources, in this case AXA customers and e-commerce platforms, is that they typically look quite different. For example, e-commerce images usually show the product in front of a white background and in ideal illumination conditions, whereas customer images contain more cluttered background and were acquired under natural illumination. As a consequence, simply training a deep network on e-commerce data to apply it to customer images is unlikely to yield accurate predictions.
To address this issue, the Computer Vision Laboratory (CVLAB) will make use of Domain Adaptation techniques that explicitly tackle this domain shift between two data sources. They will adapt this approach to the problem of price prediction from customer images. This methodology can also contribute to the evaluation of damage to a vehicle.
This research project is led by the CVLAB, of professor Pascal Fua. It lasts 6 months and is sponsored by AXA.