Abstract: A method of training a neural network for home improvement visualization is disclosed. The method may include receiving a set of annotated images of houses to train the neural network, augmenting the set of annotated images of houses to generate additional annotated images of houses for the neural network, processing the annotates images of houses and the additional annotated images of houses to train the neural network, refining the set of annotated images of houses by detecting the edges elements in the images and adjusting the annotations based on the edge detection, and outputting a trained neural network.