Using AI, the work addresses the shortage of training data, paving the way for a new era of deep learning-assisted urban landscaping.
Artificial intelligence (AI) and deep learning are everywhere, and now they hold the potential to reshape urban landscapes. Deep-learning models analyzing landscape images can aid urban planners in visualizing redevelopment plans to improve aesthetics and prevent costly errors. However, for these models to be effective, they need to accurately identify and categorize elements within images, a challenge known as instance segmentation. This challenge is due to a scarcity of suitable training data, as generating accurate “ground truth” image labels involves labor-intensive manual segmentation. However, a recently published paper shows that one team may have the answer.
Innovative synthetic data generation via AI
Researchers at Osaka University have devised a solution to this problem by leveraging computer simulation based on AI to train data-hungry models. Their approach involves creating a realistic 3D model of a city to generate ground truth segmentation. Then, an image-to-image model generates photorealistic images based on the ground truth data. This process results in a dataset of lifelike images resembling actual cities, complete with precisely generated ground-truth labels, eliminating the need for manual segmentation.
While synthetic data has previously been used in deep learning, their approach differs by creating adequate training data for real-world models through city structure simulation. By procedurally generating a 3D model of a realistic city and using a game engine to create segmentation images, they can train a generative adversarial network to transform shapes into images with realistic city textures, yielding street-view images.
This approach eliminates the necessity for publicly unavailable datasets of actual buildings and enables the separation of individual objects, even when they overlap in images. It significantly reduces human effort while producing high-quality training data. To validate its effectiveness, the researchers trained a segmentation model on simulated data and compared it to one trained on real data. The AI models performed similarly on instances involving large, distinct buildings, with a massive reduction in dataset preparation time.
The researchers aim to enhance the image-to-image model’s performance under diverse conditions. Their achievement not only addresses the shortage of training data but also lowers the costs associated with dataset preparation, paving the way for a new era of deep learning-assisted urban landscaping.