TREE AND FOREST MODELING WITH LATENT REPRESENTATION

Vegetation is ubiquitous in almost all environments, ranging from vast outdoor landscapes to indoor spaces. Consequently, models of trees serve as essential assets in applications such as games and movies, mixed reality, architecture and urban planning, agriculture, and forestry, or even the training of autonomous agents. Our research focused on high-quality tree model generation with computer graphics techniques and empowering the modeling process with artificial intelligence. We built a new generative model with deep neural network for tree-situated latent representation which makes predictions for branch growth based on where a branch node is located in the growth space of a tree model. Besides, tree model reconstruction from Lidar and drone scanning point clouds is also an open research topic, and forest reconstruction is more challenging. We developed an info-rich lookup dataset gallery generated from synthetic data and trained a neural network to encode the latent space. We segmented the forest into tree parts and replaced them with tree parts in the lookup dataset by selecting the closest neighbor in the latent domain. Our models can significantly decrease the complexity and time consumption of model generation compared with SOTA procedural modeling and generate high-quality tree models in a variety of species, and the reconstruction pipeline outperforms the SOTA in single tree reconstruction and show robustness in forest reconstruction.