Procedural System for Urban Forest Generation : Image-based Natural Terrain Generation

Sammanfattning

Procedural content generation is a common way to create model resources in the computer graphics area. It is beneficial for generating large-scale models such as buildings, forests, etc. This project focus on the generation of the urban forest, which is a case often overlooked. The placement of trees is vital for the realism of a procedurally generated scene. After analyzing the advantages and limits of current studies, the project proposes a novel system with a placement algorithm for generating forests in city areas. The system first takes a satellite image as the input and outputs a marked image. Then a placement algorithm is designed specifically for generating trees in the marked part of the image. To make the distribution more realistic, the placement algorithm combines regularity and scalable randomness with adjustable parameters including jittered range, rejection number, and cluster ratio. The system is validated by a perceptual user study and quantitative evaluation. In both experiments, the designed placement algorithm is compared with 3 other baseline algorithms (regular, random, and boundary algorithm) in 4 different map layouts. The user study is a perceptual experiment that utilized 2-AFC methods to test the effectiveness of the system and investigate how the parameters contribute to realism. The results come from 30 participants in total. The quantitative evaluation compared the distance between certain algorithms and a manually-crafted scene based on real images. The results show that randomness (controller by the parameter jittered range) is the top factor that contributes to the perceived realism of urban forests. Participants also tend to base their perception on the specific map layout. Besides, the clustering of trees (controller by the parameter cluster ratio) also increases the realism of the scene. The system is able to efficiently generate urban forests, but the realism of the results depends on the map and parameters of the algorithm.