AGRARIAN/greek_sheep_goats_dataset

AGRARIAN-EU: Greek Sheep and Goats (GSG) Object Detection Dataset

This dataset was developed within the framework of the AGRARIAN-EU project. It aims to provide high-quality training data for the detection of sheep and goats in realistic, diverse, and challenging agricultural environments.

📊 Dataset Structure

The dataset comprises 3,600 images (640x640 resolution), generated by partitioning 450 high-definition (1920x1080) video frames into 8 partially overlapping patches.

Split Statistics

Split	Image Count	Resolution	Temporal Assignment
Training	2,688	640x640	First 75% of frames per video
Validation	912	640x640	Final 25% of frames per video
Total	3,600	640x640	450 Source Frames

Key Features

• Manual Annotation: Every image has been manually labeled to ensure high-fidelity ground truth for model ingestion. Label Studio was used for the annotation process. The image below shows the manual annotation of a frame containing a herd of about 250 goats (no sheep here).
• Variable Density: Scenes range from empty landscapes to high-density clusters containing hundreds of animals.
• Leakage Prevention: To ensure valid performance metrics, the training/validation split is temporal and per-video. No frames in the validation set occur between frames used in the training set; they always follow chronologically. This ensures the model generalizes to new environmental conditions within the same flight.

---

🛰️ Data Collection Methodology

Acquisition Parameters

The raw data was collected in Chania, Greece, using a DJI Mavic 3M drone. The RGB camera captured 1080p, (i.e, FullHD, 1920x1080) footage across various flights specifically planned to maximize data diversity.

• Altitude: Variable flight heights ranging from 10m to 60m above ground level.
• Species Representation: Balanced representation of both Sheep and Goats.

Environmental Diversity & Realism

Unlike standard datasets that often feature high-contrast subjects (white sheep) grazing on uniform green grassy fields, this dataset focuses on "true-to-life" farming conditions (at least, in Greece):

• Complex Terrains: Includes rocky mountainous slopes, dirt paths, unpaved roads, and traditional pastures.
• Occlusions: Natural occlusions (tall trees and vegetation), as well as man-made structures (walls, shelters).
• Distractors: Presence of non-target objects such as scattered rocks, vehicles, tires, feeding stations and fencing.