Supervised multi-modal dictionary learning for clothing representation

Clothing appearances have complex visual properties, such as color, texture, shape and structure. Different modalities of visual features provide information complementary to each other. Combining multi-modal visual features can lead to a comprehensive description of Clothing appearances. Meanwhile, categories provide sufficient semantic information, which can lead to discriminative representations. Clothing categories exhibit hierarchical structure, which could benefit the learning algorithm. In this paper, we propose a multi-view learning algorithm, named Supervised Multi-modal Dictionary Learning (SMMDL), which learns a latent space encoding multi-modal visual properties and semantic relationships between clothing samples. Experiments on the image classification task show that SMMDL outperforms baseline methods.