Visual Clothing Recommendation

Photographs of clothing items have unique visual characteristics based on the type of color, prints/patterns, style etc. A classic problem in this scenario is the retrieval of similar clothing given a query clothing image. A re-formulation of this problem results in a co-ordination recommendation algorithm. Specifically, given a query clothing is it possible to recommend clothing that "go well" with a given query clothing? We answer this question by learning the notion of clothing items "going well" in a data driven manner using a large dataset of fashionistas wearing clothing items which are fashionable and trendy. We specifically utilize color based cues, and observe that the notion of clothing items going well with one another can be learnt based on context such as place or season from which the query is made. This problem reduces to a standard collaborative filtering formulation which we borrow to solve the clothing recommendation task.

Illustration of context based recommendation. Different tops are recommended for the same pair of blue jeans.

Top 3 retrieved items from “tops", recommended by each algorithm for query “skirts". Rows correspond to a query from a given pattern. Recommendations by TAR are more preferred than those by PR and MCL. Recommendations by CNNC and GMM have more patterns, even for patterned queries. Note that code words and modes in GMM yield identical retrievals for multiple queries.

Related Publications

Large Scale Visual Recommendations From Street Fashion Images 
Vignesh Jagadeesh, Robinson Piramuthu, Anurag Bhardwaj, Wei Di, Neel Sundaresan 
Knowledge and Data Discovery, KDD 2014.