plantnet/PlantNet-300K

Provides 306K real-world plant images across 1,081 species with carefully curated train/val/test splits and JSON metadata files mapping images to species, enabling evaluation on realistic class-imbalance scenarios. Includes PyTorch training code and pre-trained ResNet weights with utilities for model loading and fine-tuning on downstream tasks. Benchmarks standard CNN architectures against the dataset's dual challenges of label ambiguity and extreme long-tail distribution, where 80% of species comprise only 11% of images.

246 stars. No commits in the last 6 months.