Using Additional Information to Improve Classification of Clothing Item

Abstract

Clothing classification can be a difficult task for clothing e-commerce sites. When it comes to distinguishing between two types of garments that look alike, discrimination of similar categories of clothing can be complicated. Even simply glancing at photos, categories like suit and tunic are difficult to interpret. Investigation of the effect of using additional information derived from photos shows that pose along with unsupervised latent space embedding (auto-encoders) improves the classification performance significantly. Validated on data obtained from a clothing e-commerce site, additional constraints provide better performance with similar model complexities. Especially human pose representing wearers gait (or advertisement attitude) increases the classification performance by 1% in accuracy.