Digitised Herbarium Image Segmentation Dataset

This dataset was created to support the training and evaluation of semantic segmentation models for plant region extraction in digitised herbarium specimens. It was created in the context of the e-col+ project (ANR‐21‐ESRE‐005).
ContentSegmentation dataset:This dataset consists of digitised herbarium specimen images and their corresponding segmented versions, where only plant regions are preserved and background elements are removed.
The archive segmentation_dataset.zip contains two main components:

train/: 2,952 images.unsegmented_images/: 1,476 original RGB herbarium images.segmented_images/: 1,476 segmented versions of the same images (plant-only regions).test/: 333 additional herbarium images, used for evaluating segmentation models on unseen data.Training and validation images were selected from the Herbarium Segmentation Dataset (https://doi.org/10.6084/m9.figshare.27685914.v1), while the test set consists of newly and manually annotated images.
Out-of-Distribution evaluation dataset:The Out Of Distribution (OOD) dataset comprises 171 unannotated images, selected to represent visually challenging conditions, including:
noisy or highly textured backgrounds,colored backdrops (e.g., yellow, pink, dark grey),intricate plant morphologies,and common digitisation artifacts like pins, overlapping components, and mosaic-like patterns.