IAS-160k-D

Dataset description

IAS-160k-D is an object detection dataset for invasive alien species. It contains 5,537 images covering 18 invasive alien animal species, with 6,031 manually annotated bounding boxes. Bounding boxes were annotated using LabelImg and converted to the YOLO format.

The dataset is divided into training, validation, and test subsets at an approximate ratio of 70:15:15:

Split	Images	Bounding boxes
Training	3,868	4,242
Validation	821	886
Test	848	903
Total	5,537	6,031

IAS-160k-D includes species from five broad animal groups: molluscs, reptiles, insects, fish, and amphibians. It provides a benchmark for training and evaluating models for invasive alien species detection. The annotations can be used with YOLO-series object detectors after configuring the dataset paths and class names.

The dataset follows a standard YOLO directory structure, with the following components:

• train/: Contains images/ and labels/ directories for training images and their corresponding YOLO-format annotation files.
• val/: Uses the same structure for validation during model development and hyperparameter selection.
• test/: Contains images and annotations reserved for final model evaluation.
• classes.txt: Lists the 18 species names in class-ID order, starting from zero. Each annotation line contains class_id x_center y_center width height, with coordinates normalized to image dimensions.
• source.json: Preserves source, license, and attribution metadata from the parent classification dataset. It contains 162,698 records, including matching records for all 5,537 detection images. Users should consult the corresponding per-image records and comply with the original image licenses.

ScienceDB mirror

This repository is a mirror of IAS-160k-D, available on ScienceDB at DOI: 10.57760/sciencedb.32371.

Usage

For YOLO training, download and extract the original ZIP archive. The supplied data.yaml template specifies the split locations and class names; set its path to the absolute path of the extracted dataset directory before training.

The Parquet files support online image browsing and loading with the Hugging Face datasets library:

pip install datasets pillow

from datasets import load_dataset

ds = load_dataset("hf-gx/IAS-160K-D")
sample = ds["train"][0]
image = sample["image"]
class_ids = sample["objects"]["category"]
boxes = sample["objects"]["bbox"]

objects.bbox uses COCO pixel coordinates [x_min, y_min, width, height]. objects.bbox_yolo retains the original normalized YOLO coordinates [x_center, y_center, width, height], and annotation_yolo preserves the original annotation text.

For streaming access without downloading the complete dataset:

train = load_dataset("hf-gx/IAS-160K-D", split="train", streaming=True)
sample = next(iter(train))

The available splits are train, validation, and test; validation corresponds to the original val split. The Parquet version preserves the original image bytes, class IDs, and split membership, without resizing or re-encoding images.

Images retain their original licenses. Author-created annotations use CC BY-NC 4.0; see LICENSE.md.