Â鶹ÒùÔº

Researchers at The University of Osaka have developed a computer graphics (CG) model, NeuraLeaf, capable of representing a wide variety of plant species and their deformations using a single, unified model. This innovative approach leverages deep learning to overcome the limitations of traditional manual modeling, opening doors for advancements in agriculture, plant science, and breeding.

Creating realistic CG models of leaves has always been challenging. Plant leaves exhibit remarkable diversity in shape and frequently undergo deformations due to growth, environmental factors, or disease. Traditional methods often required manual creation of individual models for each species and deformation, a time-consuming and labor-intensive process.

This new method utilizes deep learning, trained on a combination of existing 2D leaf image datasets and a newly acquired 3D dataset capturing various leaf deformations. NeuraLeaf disentangles the base shape of a leaf, which varies between species, from its 3D deformations, such as wilting or curling. This allows the model to accurately represent both the species-specific characteristics and dynamic changes in leaf shape using distinct parameters.

The ability to accurately capture and track detailed changes in leaf shape has significant implications for agriculture. By fitting the NeuraLeaf model to real-world observations, researchers can monitor the growth and health of individual plants with unprecedented precision. This has the potential to improve growth prediction, enable early disease detection, and optimize resource management in agricultural practices. Furthermore, NeuraLeaf could become a valuable tool in plant breeding and scientific research.

Dr. Fumio Okura, who led the research, states, "This work is part of our 'PlantTwin' project, aimed at creating digital twins of plants. We believe this technology will revolutionize agriculture and plant science by enabling growth simulation, breeding evaluation, and a deeper understanding of plant morphology."

The findings are on the arXiv preprint server, and this research has been accepted as a highlight paper at IEEE/CVF International Conference on Computer Vision () 2025.