Assessment of photos in albums based on aesthetics and context
Dmitry Kuzovkin · HAL (Le Centre pour la Communication Scientifique Directe) · 2019
An automatic photo assessment can significantly aid the process of photo selection within photo collections. However, existing computational methods approach this problem in an independent manner, by evaluating each image apart from other images in a photo album. In this thesis, we explore the modeling of photo context via a clustering approach for photo collections and the possibility of applying such context information in photo assessment. To better understand user actions within photo albums, we conduct experimental user studies, where we study how users cluster and select photos in photo collections. We estimate the level of agreement between users and investigate how the context, defined by similar photos in corresponding clusters, influences their decisions. After studying the nature of user decisions, we propose a computational approach to model user behavior. First, we introduce a hierarchical clustering method, which allows to group similar photos according to a multi-level similarity structure, based on visual descriptors. Then, the photo context information is extracted from the obtained cluster data and used to adapt a pre-computed independent photo score, using the statistics-based data and a machine learning approach. In addition, as the majority of recent methods for photo assessment are based on convolutional neural networks, we explore and visualize the aesthetic characteristics learned by such methods.