Synthetic dataset for Sheep vs predator
Description
Description This dataset provides a large-scale synthetic UAV (unmanned aerial vehicle) imagery corpus designed to support research in automated livestock protection, specifically the detection of predators threatening grazing sheep flocks from aerial viewpoints. Real-world UAV footage of predator–livestock encounters is rare, dangerous to capture, and expensive to annotate; this dataset addresses that scarcity by synthesizing photorealistic aerial scenes with pixel-accurate annotations derived automatically from the generation pipeline. Generation Methodology. The dataset was produced using an Ecology-Aware Synthetic Data Pipeline (EA-SDP) that composes real UAV backgrounds, real sheep instances, and real predator instances under semantic terrain constraints. Backgrounds are drawn from the AeroScapes dataset (3269 semantic segmentation images, 720×1280 px). Sheep instances are extracted from the Aerial Sheep dataset (Roboflow, 4133 annotated images) via Otsu-based foreground segmentation inside YOLO-format bounding boxes. Predator instances are derived from five curated seeds (tiger, bear, crocodile, coyote, feral dog) using HSV thresholding and morphological refinement. To ensure ecological plausibility, sheep are placed exclusively on vegetation regions and predators on vegetation or construction regions, verified using a corner-and-center terrain check that requires all bounding-box sample points to fall on valid terrain. Predators are constrained to remain at least 200 pixels from every sheep per frame through a velocity-reflection mechanism. Photorealism is achieved through a four-stage compositing pipeline: color matching, edge feathering, shadow synthesis (shifted Gaussian-blurred mask multiply), and additive noise injection. Contents. The dataset comprises 1000 video clips with 100–150 frames each, yielding approximately 120,000 fully annotated frames. Each frame is stored as a 720×1280 JPEG image with a paired YOLO-format .txt annotation file. Two classes are labelled: class 0 (sheep) and class 1 (predator). Video-level train/validation splits are provided to prevent temporal leakage during model training. Verification. Ecological compliance exceeds 95% for sheep and 98% for predators. Per-frame predator–sheep separation remains strictly above the 200-pixel threshold. Class size distributions and distance statistics are provided in the accompanying dataset_statistics.json. Intended Use. The dataset supports training and benchmarking of object detection models (YOLO family, Faster R-CNN), semantic segmentation models (U-Net, DeepLab), and multi-object tracking algorithms for UAV-based precision livestock farming. It is also intended as a reproducible baseline for synthetic-to-real transfer studies in wildlife monitoring and human–wildlife conflict mitigation. Citation. If you use this dataset, please cite the accompanying paper (reference to be added upon publication). License. CC BY 4.0.