Abstract

The objective of this machine learning experiment is to generate a new map using a pix2pix Generative Adversarial Network (GAN) for image to image translation. GAN network architecture simultaneously uses a discriminator and a generator, which compete against each other. One takes images from the dataset and determines if it is a fake (generated) or real image, while the other generates new images, respectively. 

Three models were tested using different hyperparameters to see how the results were affected. One hundred images and ten epochs were used in the first trial. The training was stable with neither the generator nor discriminator dominating. A thousand iterations were produced in approximately thirty minutes. Some elements were missing and the results were not well defined. It was concluded that either the dataset or the number of epochs was not sufficient to produce a reasonably good outcome. 

The same number of images were used in the second trail, but the number of epochs was increased from ten to thirty. The model remained stable once again with neither the generator nor discriminator dominating. Three thousand iterations were generated in approximately three and a half hours. The definition of the streets and blocks in the map improved. However, the main road was still not well defined. 

The output in the third trial increased proportionally with the increase in the dataset size. The dataset size was three hundred images and the epochs were set to ten. The model remained stable, generating three thousand iterations in about three and a half hours. All the elements were visible and well defined in the final model. However, there is still room for further improvements by increasing the number of epochs. It may also be worth adjusting the layers of the network to see if this might improve the images even more. The conclusion was that the pix2pix model can generate satisfactory results with a relatively small dataset (300 in this case), and adjusting the hyperparameters may generate an even better result.

Credits

Pix2Pix Maps is a project of IAAC, Institute for Advanced Architecture of Catalonia developed in the Master in Advanced Computation for Architecture & Design in 2020/21 by:

Students: Shelley Livingston and Alexander Tong  Faculty: Stanislas Chaillou and Oana Taut