Towards Generalization of Models on Streets Imagery: Methods and Applications

Abstract

The domains relevant to urban planning have been disrupted by the proliferation of highly granular city data and the advancements in machine learning. However, machine learning models are susceptible to pitfalls constraining their deployment in many applications including domains related to urban settings. There is much to be addressed between the methods and applications before we can realize all potentials of machine learning to improve urban life. In this thesis, we focus on the use of streets imagery and classification problems. We start motivating the thesis with a case study where deep learning models are trained to predict street contexts (i.e. residential, park, commercial...etc) from streets imagery. We then shift gears and discuss a novel unsupervised domain adaptation method to address the drop in accuracy when models are tested outside the domain of the training data (i.e. a model trained on San Francisco and tested in Boston). We further our discussion with a proof of concept of a framework to develop more generalized models starting with a prototype of a system of streets imagery collection, labeling, and ending with how we approach generalization by breaking the problem into smaller prediction tasks to aid in more understanding of the interworking of the models.