Google Satelite Image to Streetmap Image translation using Pix2Pix GAN
The Pix2Pix GAN is a general approach for image-to-image translation. It is based on the conditional generative adversarial network, where a target image is generated, that is conditioned on a given input image. The idea of Pix2Pix GAN was proposed in this paper. According to the paper, the model not only learn the mapping from source image to target image, but also learn a loss function to train this mapping.
Network Architecture
The generator is a modified U-net model, it takes an RGB image as input and then tries to map it to another RGB image of the same shape. The discriminator is a PatchGan and which outputs a 30x30 matrix, which is then used to calculate the adversarial loss.
Dataset
The dataset can be downloaded from Kaggle by using this link. After downloading the dataset extract it to the data/dataset folder.
Hyperparameters
source_images = 1096
target_images = 1096
IMAGE_HEIGHT = 256
IMAGE_WIDTH = 256
IMAGE_CHANNEL = 3
DISCRIMINATOR_LEARNING_RATE = 0.0002
GENERATOR_LEARNING_RATE = 0.0002
BATCH_SIZE = 1
EPOCHS = 180
BETA1 = 0.5
BETA2 = 0.999
WEIGHT_INIT_STDDEV = 0.02
Results
Acknowledgment
This playlist helped me a lot.