Learning the Latent "Look": Unsupervised Discovery of a Style-Coherent Embedding from Fashion Images

摘要: What defines a visual style? Fashion styles emerge organically from how people assemble outfits of clothing, making them difficult to pin down with computational model. Low-level similarity can be too specific detect stylistically similar images, while manually crafted style categories abstract capture subtle differences. We propose an unsupervised approach learn style-coherent representation. Our method leverages probabilistic polylingual topic models based on attributes discover set latent factors. Given collection unlabeled fashion our mines for the styles, then summarizes by they mix those styles. organize galleries without requiring any labels. Experiments over 100K images demonstrate its promise retrieving, mixing, and summarizing their style.