In this study, we propose an efficient city-generation method based on user sketches. The proposed framework combines Conditional Generative Adversarial Networks(cGAN) and procedural modeling, which we call the Neurosymbolic Model. For cGAN training, the data set needs to consist of linked input and output pairs, so first the building of random height is generated using Perlin noise as the training data set. Then, the building contours are extracted by morphological transformation. For training, we use pairs of height maps created from the city data and sketches extracted by morphological transformation. Allowing users to generate diverse and satisfying cities from freehand sketches.
|