Pl@ntNet-300K: Plant Species Images for Set-Valued Classification

306,146 images of plants covering 1,081 species, built from the Pl@ntNet citizen observatory database. The dataset exhibits strong class imbalance and contains many visually similar species, making it suitable for evaluating set-valued classification methods. It was created by Camille Garcin of the Centre National de la Recherche Scientifique and presented at NEURIPS 2022.

Use Cases

• Benchmarking set-valued classification methods based on the described class imbalance and visual similarity.
• Training models for fine-grained plant identification based on the 1,081 species.
• Developing techniques to handle long-tailed distributions in image datasets based on the described imbalance.
• Evaluating model performance using the provided top-K and average-K metrics.

Strengths

• 306,146 images provide a substantial volume of visual data.
• Covers 1,081 distinct plant species, offering a wide taxonomic scope.
• Includes baseline results from a ResNet50 model trained with cross-entropy loss.
• Provides utilities for loading data and training models with PyTorch.

Limitations

• Exhibits a strong class imbalance, meaning a few species dominate the image count.
• Column-level documentation is absent; field semantics must be inferred after download.
• Many species are visually similar, which may complicate model training and evaluation.

Provenance

Source

Pl@ntNet citizen observatory database.

Collection Method

Built from citizen science image contributions.