Building recognition with computer vision

In this thesis we present the design, implementation and deployment of a building recognition system. First, we have developed a pipeline to find and retrieve relevant images containing landmarks and buildings from the Internet. Moreover, we propose an unsupervised approach to clean the noise present in the classes associated with each landmark/building. This method for mining and denoising images has been presented in the WebVision Workshop at CVPR'18. Then, we introduce the full system for the building recognition module, which has been developed using different computer vision and deep learning techniques. Finally, we propose a set of experiments that show the difficulties of working with a noisy dataset, we analyze the performance of the networks based blocks and finally we show the global results of the building recognition system.