HelloPlant/Orchid2024_raw

Dataset Card for Orchid2024

The Orchid2024 dataset is a fine-grained classification dataset specifically designed for cultivars of Chinese Cymbidium orchids (Chinese orchids). The dataset's samples come from 20 cities across 12 provincial-level administrative regions in China, covering 1,269 cultivars from 8 Cymbidium species, and 6 additional categories, totaling 156,630 images. The dataset nearly includes all common Chinese orchid cultivars currently found in China. Its fine-grained nature and focus on real-world scenarios make it a unique and practical resource for scientific research and real-world applications.

The dataset contained in the current repository is a full version of Orchid2024 data. The dataset authors achieved the highest Top-1 accuracy of 86.14% and a Top-5 accuracy of 95.44% on this dataset.

Dataset Features

• Most existing flower or fine-grained datasets are only aimed at species-level identification (such as distinguishing different types of orchids), while Orchid2024 extends fine-grained classification to the cultivar level, i.e., detecting different cultivars within orchid species.
• Due to the high ornamental value of Chinese orchids and the large differences in value between cultivars, the dataset exhibits a significant long-tail distribution. The number of images for each cultivar in this dataset ranges from 10 to 1387, which is basically consistent with their market popularity: common cultivars have abundant images, while rare cultivars have relatively fewer.
• The Orchid2024 dataset features a two-level fine-grained classification system of species and cultivars, which can accurately identify subtle visual differences between different cultivars. This dataset covers 8 orchid species and their subordinate cultivars, some of which have significant differences, while others have similar forms; and within the same species, the difficulty of distinguishing cultivars is even higher.

Dataset Format and Structure

Data Format

The Orchid2024 dataset contains three parts: training images, training information files, and data description files. The training images are divided into training, validation, and test sets in a ratio of 6:2:2, specifically including 94,036 training samples, 31,297 validation samples, and 31,297 test samples. Each set is organized by category in the form of subfolders, covering a total of 1,275 categories, with each subfolder corresponding to a category, which stores JPG format images of the same category.

Dataset Loading Method

You can load the specified data yourself according to actual needs. The loading method example is as follows:

from datasets import load_dataset

dataset = load_dataset("HelloPlant/Orchid2024")
# check dataset splits
print(dataset.keys())
# View the first sample in the train set
print(dataset["train"][0])
print(f"Dataset size: {len(dataset['train'])}")

# Iterate through each image in the test dataset
for data in dataset['test']:
    # Get training data
    img = data['Img'] # Get image data
    label = data['Label'] # Category label ID corresponding to the image
    # Get attribute data attached to the image
    class_name = data['Class_Name'] # Get class name
    cultivar_name = data['Cultivar_Name'] # Get cultivar name (English)
    species_name = data['Species_Name'] # Get species name (English)
    Chinese_cultivar_name = data['Chinese_Cultivar_Name'] # Get cultivar name (Chinese)
    Chinese_species_name = data['Chinese_Species_Name'] # Get species name (Chinese)
    # Perform subsequent processing...
    break  # Only display the first piece of data

Citation Information

If this dataset is helpful to you, please cite the following related paper:

@article{Peng2024Orchid,
  title={Orchid2024: A cultivar-level dataset and methodology for fine-grained classification of Chinese Cymbidium Orchids},
  author={Peng, Yingshu and Zhou, Yuxia and Zhang, Li and Fu, Hongyan and Tang, Guimei and Huang, Guolin and Li, Weidong},
  journal={Plant Methods},
  volume={20},
  number={1},
  pages={124},
  year={2024},
}