Neural Hierarchical Decomposition for Single Image Plant Modeling

2 RIKEN AIP 3 SIAT, Chinese Academy of Sciences Figure 1 . We propose a novel learning-based framework that automatically generates high-quality 3D plant models from single-view images. Our method sequentially combines hierarchical structure learning with parametric modeling based on botanical priors, producing usable, structured plant assets ready for immediate use in practical applications. Through learning the decomposition of box hierarchies at different levels of detail (LoD), our method offers a comprehensive solution that can adapt to two prominent categories of plants: (a) Houseplants and (b) larger trees. (c) shows a garden with several of our results directly assembled.