Online eCommerce is characterised by the problem of presented many visually similar and relevant items to users before finalising their purchase. Traditional recommender systems operate on collaborative filtering techniques don’t factor in image content or visual details, as they rely on user clicks / past purchases. Visual recommendation engines need to go beyond the old school visual attributes such as color, shape or texture similarities. They need to take into account the semantics of the product and its attributes, and retrieve Visually similar items from the catalog.
Our engine deals with visual similarity searches not only from the “catalog” but also “from the wild” (user uploaded similar images), by quantitatively estimating visual similarity, applying higher levels of product abstractions beyond just colors and patterns.
For example, if an user in interested in a particular type of lamp in a a living room, the search can be focused on just that object within that region of interest (RoI), and visually similar objects can be presented for further browsing and purchasing.