Visual Siamese Clustering for Cosmetic Product Recommendation

Abstract

We investigate the problem of a visual similarity-based recommender system, where cosmetic products are recommended based on the preferences of people who share similarity of visual features. In this work we train a Siamese convolutional neural network, using our own dataset of cropped eye regions from images of 91 female subjects, such that it learns to output feature vectors that place images of the same subject close together in high-dimensional space. We evaluate the trained network based on its ability to correctly identify existing subjects from unseen images, and then assess its capability to find visually similar matches amongst the existing subjects when an image of a new subject is input.