LeafSnap30 is a (modified) subset of the images from the 30 species with the highest number of images from the LeafSnap dataset. LeafSnap is an electronic field guide for identifying tree species from photos of their leaves. The original dataset consists of images taken from 2 different sources as well as their segmented versions using the LeafSnap segmentation algorithm. The two sources are: high quality "lab" images of pressed leaves from the Smithsonian collection and "field" images taken by mobile devices in outdoor environments. The original "lab" images contain size and color calibration rulers, which interfere with the training of end-to-end Deep Learning (DL) models for automatic tree species classification. Therefore, we have semi-manually cropped the "lab" images of the 30 species with most number of images in order to keep only the leaves and we do not include the segmentation masks. The original "lab" leaf images are also included in the dataset, but the file paths point only to the cropped ones. The original dataset has been released in 2012 (before the DL revolution in Computer Vision) in order to promote further research in leaf recognition. The authors ask their paper to be sited (see original link above) if the dataset is used. We are releasing the cropped subset as the LeafSnap30 dataset in order to demonstrate the performance of eXplainable AI (XAI) methods applied on DL models trained to solve simple, yet realistic scientific problem.

This dataset will be used for demonstration purposes in the open-source Deep Insight and Neural Network Analysis (DIANNA) project, whose goal is to provide a library for explainable AI methods for scientists. DIANNA is work in progress at the time of publishing this dataset version (July 2021): https://github.com/dianna-ai/