LAESI: Leaf Area Estimation with Synthetic Imagery

TL;DR

The paper tackles the scarcity and cost of real annotated leaf data by introducing LAESI, a synthetic dataset of 100K leaf images rendered on millimeter paper with precise semantic masks and leaf-area labels. It combines fast, controllable 3D procedural leaf and background generation in Unity with a ControlNet-based inpainting pipeline and a filtering step to ensure annotation consistency, enabling large-scale, domain-relevant data generation. Validation shows that models trained with LAESI data can achieve leaf-area predictions with a mean relative error competitive with or better than human annotators, while preserving segmentation performance, indicating strong potential for agriculture and biology applications. The work demonstrates that synthetic data, when carefully filtered and inpainted, can significantly improve domain-specific vision tasks and can be adopted for remote sensing and precision agriculture pipelines.

Abstract

We introduce LAESI, a Synthetic Leaf Dataset of 100,000 synthetic leaf images on millimeter paper, each with semantic masks and surface area labels. This dataset provides a resource for leaf morphology analysis primarily aimed at beech and oak leaves. We evaluate the applicability of the dataset by training machine learning models for leaf surface area prediction and semantic segmentation, using real images for validation. Our validation shows that these models can be trained to predict leaf surface area with a relative error not greater than an average human annotator. LAESI also provides an efficient framework based on 3D procedural models and generative AI for the large-scale, controllable generation of data with potential further applications in agriculture and biology. We evaluate the inclusion of generative AI in our procedural data generation pipeline and show how data filtering based on annotation consistency results in datasets which allow training the highest performing vision models.

Figures (12)

• Figure 1: A model of the LAESI pipeline: Procedural Generation of Millimeter Paper Background and Leaf Shape generates diverse paper textures, grid alignments, and a range of leaf shapes, sizes, and textures. Rendering and Final Synthetic Dataset Composition combines leaves with the background with realistic lighting and generates annotations such as semantic masks, surface area labels, and canny edges. Dataset Inpainting utilizing the ControlNet-based pipeline for inpainting of Canny edges generates leaf images inside the masked regions of data points. Dataset Filtering discards the leaf data points with inpainting results that reduce consistency with their annotations by using a semantic segmentation model.
• Figure 2: Selection of different millimeter papers generated using our procedural shader method ranging from sharp to blurry.
• Figure 3: Procedural leaf model generation: The shape is defined by a parametric curve using Unity's Animation Curve (a), which is then textured, including vein pattern development (b), and stochastic elements and surface details are added via shader effects (c).
• Figure 4: Diverse final renderings from the procedural leaf generation pipeline. This collection illustrates the variation achieved in leaf appearance through our procedural model parameters. Each rendering captures different lighting conditions, shadow effects, and background scaling.
• Figure 5: Three instances of ControlNet-generated images where the region of the inpainted leaf in the mask deviates significantly from the region defined by the procedurally generated mask. Such data points are automatically filtered out in LAESI.