Skip to content

🎯 Few-shot Learning

Overview

Rare species are the norm, not the exception!
The few-shot benchmark isolates species with fewer than five training samples. The task: Can your model generalize with minimal supervision?


Use Cases

  • Identification of rare, endangered, or newly observed fungi
  • General few-shot learning research (prototypical networks, embedding-based methods, etc.)

Data & Splits

  • Only species with <5 training images are included.
  • Standard temporal splits: train (up to 2021), validation (2022), test (2023).

Evaluation Protocol

  • Primary Metric: Top-1 accuracy
  • Secondary Metrics: Macro F1-score, Top-3 accuracy

Baselines & Results

Includes nearest-neighbor, centroid-prototype, and standard classifiers on strong feature extractors (CLIP, BioCLIP, DINOv2, BEiT).

See Baselines & Models for results and example code.


Quick Start

  • Few-shot split provided in the dataset package and Kaggle.