Example-based rapid generation of vegetation on terrain via CNN-based distribution learning

Abstract

Modeling large-scale vegetation on terrain is an important and challenging task in computer games, movie production and other digital entertainment applications. In this work, we propose a novel example-based method for rapid generation of vegetation in outdoor natural environments. Its central idea is to learn the vegetation distribution on terrain via deep convolution neural networks. We first use a pre-trained deep neural network to extract rich local information from the terrain pertinent to vegetation distribution. Second, we produce the initial features of the target vegetation distribution based on patch matching and further introduce a network that generates a vegetation density map based on the initial features. Third, during the synthesis stage, we propose a procedural method to generate the vegetation distribution data corresponding to the terrain data. Our research work confirms that the image features extracted by the pre-trained deep neural network could be utilized to explore the connection between vegetation and terrain. We validate our new method over various outdoor scenes, including procedural generated scenes and scenes with manual control on tree patterns. The experimental results demonstrate that our method can rapidly produce new realistic scenes for outdoor natural environments, which relies on the mechanism of learning correlationship between vegetation distribution and terrain data.

Evaluation metrics for intelligent generation of graphical game assets: a systematic survey-based framework

“…Operation evaluation includes performance and controllability measures. Performance metrics assess resources, such as memory [51], [52], [53] or time cost [29], [52], [53], [54], [55] of a method. While these metrics are important during development and testing, they are particularly important when applications go into production and release and run in real-time, as they have direct implications for userexperience and usability.…”

“…Objective similarity metrics are applied in the comparison between corresponding data points, such as mean absolute error (MAE) [17], [51], [57], mean squared error (MSE) [36], [37], [38], [39], root-mean-squared error (RMSE) [2], [3], [4], [40], [41] and sum of squared errors (SSE) [23]. MAE aggregates the absolute error of data-points, that is the positive difference between corresponding values, while SSE and MSE square these errors strengthening larger errors and diminishing smaller errors.…”

Translation of Real-World Photographs into Artistic Images via Conditional CycleGAN and StarGAN

“…Because CNNs are good at extracting important features even if the inputs are distorted, they have been applied to image processing tasks, such as image recognition (AlexNet [30]) and object detection (Faster R-CNN [31]). CNN is also utilized in creative task such as super-resolution of face images [32], generating terrain [33], and generation of conditional single text from alphabet to katakana (Japanese alphabet for transcribing foreign words) [34].…”

VR content creation and exploration with deep learning: A survey

“…An erosion meta-simulation is also trained to efficiently apply erosion to terrains in their work. Zhang et al [100] proposed an example-based method to rapidly generate vegetation in outdoor natural environments. The method utilizes a VGGnetwork to learn the relationships between terrain and vegetation distribution.…”