Title: Understanding online fashion networks

Abstract : From Instagram to Pinterest, visual media becomes one of the most important form of information in online social networks. Billions of images produce social interaction through visual media on the Web. But how can we model content and interaction in such large-scale visual networks? This dissertation research tries to approach this question using data from Chictopia, a real-world online fashion network. I consider two major problems in this context: understanding of visual content, and understanding of user behavior.

Understanding visual content is the ultimate goal in computer vision. In this dissertation, I study computer vision techniques to recognize garment items in a picture. I propose clothing-parsing algorithms, which assign one of clothing category to every pixel. The algorithm takes advantage of the unique characteristics of fashion pictures; human pose gives a strong contextual cue in clothing parsing. Using this knowledge, I first formulate clothing parsing as a joint label assignment with respect to a probability distribution, and show that this approach results in good localization performance. The second formulation also considers detection of items, using a data-driven approach. The empirical results show promising recognition performance and the benefits of clothing parsing in human pose estimation.

On understanding user behavior, I study what factors are affecting the popularity of pictures in this dissertation. Using the clothing parsing techniques as well as network and text information, I predict the number of votes in both in-network and out-of-network scenarios. The experiments find significant statistical evidence that social factors dominate the in-network scenario, but a combination of content and social factors can help predicting popularity outside of the network.