MacroSony/inat-2017-subset

iNaturalist 2017 Supercategory Subset

This dataset is a sampled subset of the iNaturalist 2017 Challenge dataset, specifically processed for efficient object detection training across major biological supercategories.

Dataset Summary

• —Source: iNaturalist 2017 (Competition Version)
• —Task: Object Detection
• —Classes: 9 (Collapsed from thousands of species into biological supercategories)
• —Training Samples: 1,000 images per supercategory (~9,000 total)
• —Validation Samples: 200 images per supercategory (~1,800 total)

Class Map (ID to Supercategory)

The original thousands of species-level categories have been remapped to the following indices:

File Structure

• —subset_train_bboxes.json: COCO-formatted annotations for the training set.
• —subset_val_bboxes.json: COCO-formatted annotations for the validation set.
• —images/: Directory containing all sampled JPEG images.

Usage & Reproduction

This subset was generated using dataprep.py by:

1. 1.Downloading the 2017 Competition bounding box annotations.
2. 2.Grouping images by their biological supercategory.
3. 3.Randomly sampling a fixed number of images per group.
4. 4.Collapsing the category IDs in the JSON files to the 0-8 range based on supercategory.
5. 5.Streaming the original 165GB iNaturalist tarball and extracting only the sampled image files.

Licensing

The images and original annotations are provided by iNaturalist and are generally under Creative Commons Attribution-NonCommercial (CC BY-NC) licenses. Please refer to the official iNaturalist 2017 page for specific license details for each image.