Saliency directed product placement

Abstract

Product placement or embedded marketing involves the process of finding strategic locations for particular objects within a scene such that the attention is shifted to the designated product. The use of advanced artificial intelligence and computer vision is expanding into every area of our lives. Recently, there has been rapid development in deep learning which aided in a number of breakthroughs in computer vision, including visual saliency. Visual saliency is the branch of computer vision that generates attention-based models of a given scene using certain image features. Through evolution, certain characteristics of an image such as colour and contrast make an object stand out from its neighbours and attract primates’ attention. These are most of the time characterised by regions of different contrast, different intensity values and even orientation of features. This project evaluates how the use of visual saliency techniques can improve product placement analysis by implementing a saliency-directed objective product ranking system. This system was developed to look at a scene with a number of products and predicts which one the customer will fixate on first. It initially uses a state-of-the-art object detector in order to detect the products in the scene and another component ranks the detected product using a technique based on a saliency segment ranking algorithm. The modular aspect of this system allows many aspects of the product ranking technique to be altered such as substituting the saliency map generating algorithm which would allow us to evaluate the effectiveness of the different configurations and which components affect the accuracy of the rankings the most. In order to determine the effectiveness of the product ranking technique, it’s product ranking results had to be compared with a true baseline ranking of the products. Spearman’s correlation Coefficient test was used to evaluate the correlation between the true baseline product ranking against the product ranking system’s output where a medium-to-strong correlation was obtained with a resultant correlation coefficient of 0.66.