基于GAN的树叶飘落模拟

Leaf floating simulation based on GAN

为了模拟在三维渲染、虚拟现实、游戏场景、电影特效等环境中具有真实感的树叶飘落效果,同时受当下日益成熟的深度学习算法的启发,提出了一种基于生成式对抗网络(GAN)的树叶飘落动态效果模拟算法。首先,通过实验或实地采集不同环境条件下(风力级别、扰动级别等)树叶飘落的6元时间序列制作训练数据集;然后使用该数据集对双通道对抗式生成网络模型(DACGAN)进行训练;其次,训练后的模型能够根据需求(风力级别、扰动级别等)输出三维落叶数据;最后,将这些数据用于各种图形渲染环境中。理想状态下,相对于传统的基于数学模型控制飘落的方法以及人为设置关键帧的动画方法,可以为捕获树叶飘落的真实感提供可能,且随着计算机视觉技术以及深度学习技术的发展成熟,使得现实中的三维数据更易获取,生成模型更加优化,DACGAN算法成本更低。

Inspired by the progress of deep learning algorithm,we propose a method based on the generative adversarial network(GAN)to simulate the effect of leaf falling,in order to achieve realistic leaf floating in the 3D rendering,virtual reality and other environments.Firstly,the training data set is produced by experiment or in the field of 6-element time series of leaf falling under different environmental conditions(wind level,disturbance level,etc.),and then the double-channel auxiliary classifier generative adversarial networks model(DACGAN)is trained using the data set.Secondly,the trained model can output 3D leaf data according to our demands(wind level,disturbance level,etc.).Finally,the data can be used in a variety of graphic rendering environments.Compared with the traditional mathematics-based method controlling the leaf floating and the artificial animation method setting keyframes,our method,as the experimental results indicate,can provide the possibility to capture the realism of the falling leaves.In addition,with the development of computer vision and deep learning,the 3D data in reality is more easily acquired,the generated model is more optimized,and the cost of DACGAN will be lower.