Research map / 3D vision / Face and shape
Neural Hierarchical Decomposition for Single Image Plant Modeling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Generates 3D plant models (trees and houseplants) from a single photo by inferring a hierarchical oriented-box decomposition with a graph VAE-style decoder, then synthesizing geometry with parametric L-system/leaf models.
Problem
Reconstructing usable, structured 3D plant models from single-view photographs, which is hard due to intricate branching and self-occlusion; existing methods handle only outdoor trees or are semi-automatic.
Approach
A neural hierarchical box decomposition (n-ary graph of oriented boxes with botanical parameters), decoded recursively from an image latent via a MolGAN-based graph decoder trained as a VAE, with level-of-detail varying between houseplants (leaf-level terminals) and trees (sub-tree terminals), followed by parametric L-system and deformable-leaf geometry generation.
Key results
- Box-structure reconstruction on synthetic test set: Chamfer distance 0.065 for houseplants and 0.082 for outdoor trees (average 0.073).
- Qualitative comparison shows Wonder3D and One-2-3-45 produce noisy, structurally inseparable volumes for trees, while the proposed method yields structured assets.
- Qualitative comparison against the RBV-based tree method (Li et al.) shows it lacks flexibility for trees with holes and long branches, and cannot handle houseplants.
Limitations
- Evaluation of real-image results is qualitative; quantitative metrics are only on synthetic data.
- Trained separately per species on synthetic data, and only 12 tree and 9 houseplant species covered.
- Botanical parameters are not learned from input images in the current version (future work).
Experimental setup
Datasets: synthetic plant dataset generated with self-organization and L-system
Benchmarks: synthetic plant test set
Metrics: Chamfer distance Hausdorff distance RMSE standard deviation
Baselines: Wonder3D One-2-3-45 Li et al. radial bounding volume method
Builds on: StructureNet MolGAN UPerNet ResNet-50 Wonder3D
Closest papers by meaning
- Learning to Infer Parameterized Representations of Plants from 3D Scans (CVPR 2026 · similarity 0.91)
- Probabilistic Plant Modeling via Multi-View Image-to-Image Translation (CVPR 2018 · similarity 0.87)
- SVDTree: Semantic Voxel Diffusion for Single Image Tree Reconstruction (CVPR 2024 · similarity 0.85)
- PlantTraitNet: An Uncertainty-Aware Multimodal Framework for Global-Scale Plant Trait Inference from Citizen Science Data (AAAI 2026 · similarity 0.84)
- BioCLIP: A Vision Foundation Model for the Tree of Life (CVPR 2024 · similarity 0.80)
- Birth and Death of a Rose (CVPR 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.