taylahmccullough/QCIF_Wildlife_Detection_Benchmarking_Dataset

QCIF Wildlife Detection Benchmarking Dataset

Dataset Description

This dataset was prepared as part of the QUT IT Capstone project, Team 13, in collaboration with QCIF Digital Research.

It was created for benchmarking wildlife detection and species classification models on Australian camera-trap images. The dataset was used to compare three models:

• SpeciesNet, developed by Google
• AWC 135, developed by Australian Wildlife Conservancy
• WildObs National, developed by QCIF Digital Research

The dataset contains 8,000 camera-trap images across eight Australian species, with 1,000 images per species. Images were sourced from the Atlas of Living Australia.

Dataset Structure

The dataset is provided as data.zip.

After extraction, the folder structure is:

data/
└── raw/
    ├── canis_familiaris/             1000 images
    ├── felis_catus/                  1000 images
    ├── hydromys_chrysogaster/        1000 images
    ├── hypsiprymnodon_moschatus/     1000 images
    ├── megapodius_reinwardt/         1000 images
    ├── perameles_nasuta/             1000 images
    ├── thylogale_stigmatica/         1000 images
    └── uromys_caudimaculatus/        1000 images

Ground truth labels are derived from folder names. No separate annotation file is required for the benchmarking workflow. For example, images stored in data/raw/felis_catus/ are treated as felis_catus.

Species Included

Folder Label	Common Name	Image Count
canis_familiaris	Domestic Dog / Dingo	1,000
felis_catus	Domestic Cat	1,000
hydromys_chrysogaster	Common Water Rat	1,000
hypsiprymnodon_moschatus	Musky Rat Kangaroo	1,000
megapodius_reinwardt	Orange-footed Scrubfowl	1,000
perameles_nasuta	Long-nosed Bandicoot	1,000
thylogale_stigmatica	Red-legged Pademelon	1,000
uromys_caudimaculatus	Giant Uromys	1,000

Usage

Download and extract the dataset:

unzip data.zip

The extracted data/raw/ folder can be used directly with the benchmarking pipeline:

https://github.com/taylahmccullough/QCIF_Wildlife_Detection_Benchmarking

Intended Use

This dataset is intended for:

• benchmarking wildlife detection and species classification models
• comparing model performance across Australian camera-trap species
• evaluating Precision, Recall, F1, false negatives, and high-confidence misclassification patterns
• supporting QCIF Digital Research handover and future model review

Out-of-Scope Use

This dataset should not be used as training data for SpeciesNet, AWC 135, or WildObs National if those same models are later evaluated against this dataset. Training on this dataset would bias future benchmark results.

Dataset Preparation Notes

The dataset was balanced to contain 1,000 images per species.

The benchmark uses folder-based labels, where the species folder name is treated as the ground truth label.

Limitations

This dataset reflects real-world camera-trap conditions. Image quality is not uniform and may include:

• partial occlusion
• poor lighting
• animals at varying distances from the camera
• unclear animal visibility
• humans, background objects, or non-target species appearing in some images

No full manual validation was performed to confirm that every image contains a clearly visible target animal. These limitations should be considered when interpreting model performance.

Licence and Attribution

Images were sourced from the Atlas of Living Australia. Attribution to the Atlas of Living Australia is required for reuse or redistribution.

Individual source records may carry additional licence terms from the original data provider. Users should verify record-level licence conditions through the Atlas of Living Australia before redistribution, publication, or downstream use.

Project Context

This dataset was created as part of the Australian Wildlife Species Detection and Monitoring project for QCIF Digital Research. It supports benchmark comparison across SpeciesNet, AWC 135, and WildObs National using a consistent eight-species Australian wildlife image dataset.