Transforming the urban landscape: The impact of artificial intelligence

AI and deep learning are already everywhere, and now they have the potential to reshape the urban landscape. Deep learning models that analyze landscape images can help city planners visualize redevelopment plans, improve aesthetics and avoid costly mistakes. However, for these models to be effective, they need to accurately identify and classify elements in images, a challenge known as instance segmentation. This challenge arises due to the lack of suitable training data, as generating accurate "ground truth" image labels involves labor-intensive manual segmentation. However, a recent paper suggests that a team may have found the answer

Generating innovative synthetic data through artificial intelligence

Researchers at Osaka University used artificial intelligence-based computer simulations to Training models that require large amounts of data, devised methods to solve this problem. Their approach involves creating a realistic 3D model of the city to generate ground truth segmentations. The image-to-image model then generates realistic images based on ground truth data. This process results in a realistic image dataset that resembles an actual city, complete with accurately generated ground truth labels, eliminating the need for manual segmentation.

While synthetic data has been used for deep learning before, their approach is different, creating enough training data for real-world models through simulations of urban structures. By procedurally generating 3D models of realistic cities and using a game engine to create segmented images, they can train a generative adversarial network to convert shapes into images with realistic urban textures, thereby generating street view images.

Advantages and future prospects

With this approach, it is no longer necessary to use publicly available datasets of actual buildings, while being able to isolate individual objects even if they overlap in the image. This approach significantly reduces labor costs while generating high-quality training data. To verify its effectiveness, the researchers trained the segmentation model on simulated data and compared it with a model trained on real data. Results showed that the AI ​​model performed similarly on instances involving large, unique buildings, but with significantly reduced dataset preparation times. The researchers aimed to improve the performance of image-to-image models under different conditions. Their achievement not only solves the shortage of training data but also reduces the costs associated with dataset preparation, paving the way for a new era of deep learning-assisted urban landscaping.

The above is the detailed content of Transforming the urban landscape: The impact of artificial intelligence. For more information, please follow other related articles on the PHP Chinese website!