GeoComp demos

Building image classification

Urban building characterization is a complex problem with many involved parts whose solution can benefit the development of smart autonomous driving systems, the digital archiving of cultural artifacts, as well as the automation of real estate valuation. To contribute to this research area, this work focuses on a specifc part of the problem: the classiffcation of urban buildings from photographs of their facades into 10 categories: Church, Mosque, Synagogue, Buddhist Temple, House, Apartment Building, Mall, Store, Restaurant, and Oce Building. For this purpose, a novel hierarchical multi-label CNN-based model is proposed. Based on the coarse-to-fine paradigm, a label tree is set up, and the model provides outputs corresponding to each level of the resulting hierarchy.
Feedback from the coarser level to the finer one runs through the model using simple probabilistic notions encoded through a multiplicative layer connecting the parent coarse branch of the model to the child ne branch. The resulting model solves the urban building classification task while performing better than both classical convolutional neural networks, as well as existing hierarchical modes.

References:

S. Taoufiq: "Urban Building Classification Using CNN-based Hierarchical Models", Master Thesis, 2020 (supervisors: Csaba Benedek, Balázs Nagy, SZTAKI)

S. Taoufiq, B. Nagy and Cs. Benedek: ”HierarchyNet: Hierarchical CNN-based Urban Building Classification,” Remote Sensing, vol. 12, no. 22, article 3794, 2020, IF: 4.848 Open Acces