PubMed · 2026-05-25
Researchers in Chile built a large, high-quality image library of 16 pollen types from the Biobío Region, complete with precise shape outlines verified by a pollen expert. This dataset is designed to train AI systems that can automatically identify pollen — a task that matters for tracking plant diversity, allergies, and honey authenticity.
The dataset contains 16,198 high-resolution microscopy images (3088×2064 pixels) with 36,383 hand-verified pollen outlines across 16 species, making it one of the most detailed pollen datasets available.
Each pollen grain was photographed at three focal depths, capturing surface and interior features invisible in single-plane images — a design that improved AI classification robustness.
A baseline AI model trained on the dataset achieved 0.985 mask mAP@50 (near-perfect detection accuracy) on the validation set after just 50 training cycles.
