Demeter: A Parametric Model of Crop Plant Morphology from the Real World

TL;DR

Demeter introduces a data-driven parametric model for crop plant morphology that jointly encodes topology, articulation, shape, and deformation to produce realistic plant meshes. By learning PCA-based shape and deformation bases from real-world field data, and by representing topology as a variable tree, Demeter enables accurate multi-view and single-view reconstruction and facilitates biophysical simulations. The approach is demonstrated on a comprehensive soybean dataset with extensions to maize, and it shows improved fidelity over existing procedural and neural reconstruction methods while offering interpretable latent factors for morphology manipulation. This work advances crop phenotyping and agronomic simulation by providing a compact, interpretable, and generalizable plant morphology model grounded in real-world data and capable of integration into biophysical models and simulations.

Abstract

Learning 3D parametric shape models of objects has gained popularity in vision and graphics and has showed broad utility in 3D reconstruction, generation, understanding, and simulation. While powerful models exist for humans and animals, equally expressive approaches for modeling plants are lacking. In this work, we present Demeter, a data-driven parametric model that encodes key factors of a plant morphology, including topology, shape, articulation, and deformation into a compact learned representation. Unlike previous parametric models, Demeter handles varying shape topology across various species and models three sources of shape variation: articulation, subcomponent shape variation, and non-rigid deformation. To advance crop plant modeling, we collected a large-scale, ground-truthed dataset from a soybean farm as a testbed. Experiments show that Demeter effectively synthesizes shapes, reconstructs structures, and simulates biophysical processes. Code and data is available at https://tianhang-cheng.github.io/Demeter/.

Figures (25)

• Figure 1: Demeter is a parametric model that is trained on real-world data. It encodes plants into shape, topology, articulation, and deformations, providing a realistic, compact representation that can generalize to many species.
• Figure 2: The Demeter Model consists of four parametric components: (a) topology $\boldsymbol{\Gamma}$, a tree-structured graph that stores the binary connection relationship between plant-part nodes; (b) articulation parameters $\boldsymbol{\theta}$ defining each node’s relative position to its parent, using quaternions for rotation and 1D translation; (c) shape parameters $\boldsymbol{\beta}$, representing leaf variance via principal component coefficients, learned from 2D leaf scans; (d) deformation parameters $\boldsymbol{\gamma}$, learned from 3D leaf point clouds and fixed 2D shape parameters $\boldsymbol{\beta}$, controlling the 3D skeletons of leaf and stem. Here the gray, orange and green nodes represents root, stem and leaf respectively.
• Figure 3: Demeter model parameters. (a) Articulation includes scale $s$, which controls the size of the component (e.g., stem length); position $d$, which controls the location of the joint node relative to its parent; and quaternion angle $\boldsymbol{\tau}$, which controls the relative rotation. (b) The mean template learned from 2D leaf scannings. (c) The shape parameter encodes the location of contour points, which defines the inner points of the leaf and serve as joints to deform the leaf and stem. (d) Deformation $\boldsymbol{\gamma}$ represents a grid of Catmull-Rom control points to manage shape and articulation-independent deformation, such as leaf curling. (e) An illustration of the mapping between the leaf shape control points and canonical UV space.
• Figure 4: PCA coefficients of leaf shape. We visualize the mean shape and the one of the principal components of leaf for several species in $(-2\sigma, 0, 2\sigma)$. For 3D deformation, we show the first and last few components for soybean.
• Figure 5: Reconstruction from single image. One2345++ generates a coarse shape but fails to capture details. Meshy reconstructs a realistic plant but is misaligned with the input image. In contrast, our model is complete and faithful to input.