Computer vision models can be used to extract information about the accessibility of a certain city from panorama images. To train such a model, we need a lot of panoramas. While this all works fairly well when using data from one city, or cities that look alike, questions remain about how such a model would perform on a city that looks visually significantly different.
We presented three experiments to examine these questions.
In the first experiment, we test a general model trained using data from North American cities on an Amsterdam test set to see if this significantly different-looking test set has an influence on the overall performance of the model.
In a second experiment, we train a model using that same Amsterdam dataset, to compare it to the results of the first experiment. This should give some interesting results, since the Amsterdam dataset is significantly smaller than the North American one. Therefore, we can examine how sensible models are to quality as opposed to quantity of data.
In a last experiment, we test whether adding some of this dissimilar data from the North American dataset could potentially improve the results of the model.
Author: Kasper Verhavert
This research was conducted by Kasper Verhavert from the Department of Sciences and Bioengineering Sciences of the Vrije Universiteit Brussel, in collaboration with the AI Team, Urban Innovation and R&D, City of Amsterdam.
Involved civil servants: Diederik M. Roijers
Supervisors: Diederik M. Roijers,
Header afbeelding: Header from Google Streetview