Abstract
Flowers are fundamentally three-dimensional structures, yet much of plant comparative work still relies on flattened specimens, limited views, or simplified descriptors. Richer 3D data could open new possibilities for studying floral form, function, and evolution, but collecting such data at scale remains difficult. Living flowers are often small, delicate, reflective, hairy, curved, or partly translucent, and even when a 3D model can be generated, turning it into biologically useful comparative data is another challenge.
We are exploring machine learning in this broader context of floral 3D phenomics. On one side, newer approaches such as neural rendering and related methods may help extend what kinds of flowers can be reconstructed from image-based datasets. On the other hand, machine-learning-based shape representations may help preserve biologically meaningful geometry for comparative and evolutionary analyses.
Together, these directions point toward a larger goal: making 3D flowers not just visually impressive, but comparable, reusable, and evolution-ready.
Pipeline Overview
This project explores image-based 3D reconstruction workflows for flowers and related questions in floral morphology. The current site highlights reconstruction results obtained so far, while methods for downstream comparative and evolutionary analysis remain under development.
Image Collection
Photographs of flowers are collected from multiple viewpoints to support image-based reconstruction.
Preprocessing
Images are organized and prepared for reconstruction, including calibration and pose estimation workflows where applicable.
3D Reconstruction
We test neural rendering and Gaussian splatting approaches to generate 3D representations of floral structures from photographs.
Qualitative Assessment
Current evaluation focuses on visual inspection of reconstruction quality, structural completeness, and preservation of visible floral features.
Results
Work is ongoing. Preliminary qualitative results from early reconstruction experiments are shown below.
Qualitative Reconstructions
Interactive 3D Gaussian-splat models — click and drag to rotate, scroll to zoom.
Evaluation
Evaluation of reconstruction quality is ongoing. At this stage, we focus on qualitative assessment of 3D reconstructions across different plant specimens, including visual fidelity, structural completeness, and the ability to capture fine morphological features.
Future work will include systematic quantitative evaluation using geometric metrics and comparisons across reconstruction methods.
Video Summary
Citation
If you find this work useful, please cite:
@misc{jin2026phyto3d,
title = {From Photos to Phenotypes: Machine Learning for Floral {3D} Reconstruction and Shape Evolution},
author = {Jin, JianJun and Eaton, Deren A. R. and Yang, Yue and Elmelegy, Raaid},
year = {2026},
note = {Project page},
url = {https://relmelegy.github.io/phyto3d/}
}