titoruizh/Drone-Orthomosaic-Vehicles-Yolo-annotation

Dataset Tailings Mining Vehicles & Instruments (High-Res Drone Imagery)

Dataset Summary

This dataset contains high-resolution aerial imagery focused on vehicle detection and geotechnical monitoring instruments within active mining environments (tailings dams). The data was acquired using a DJI Zenmuse P1 sensor at 120m altitude.

Photogrammetric Context

The images originate from large-scale georeferenced orthomosaics generated from bi-daily photogrammetry flights.

• Adaptive Tiling Process: The orthomosaics were tiled into JPG segments with specific dimensions optimized for the target object size:
  • Vehicles & Machinery: Large tiles (~7500x7500px) to maintain broad environmental context.
  • Piezometers: Smaller, granular tiles (1024x1024px) were extracted to preserve maximum resolution for these smaller geotechnical instruments.
• Note on Georeference: While the source orthomosaics were georeferenced (WGS84 / Local Mine Grid), individual JPG tiles do not retain geospatial metadata.

Dataset Structure

The dataset is organized into three distinct subsets based on the object category. Each folder contains its own images and labels subdirectories.

Dataset Tailings mining vehicles/
├── CARS-TRUCKS/
│   ├── images/
│   └── labels/
├── MACHINERY/
│   ├── images/
│   └── labels/
└── PIEZOMETER/
    ├── images/
    └── labels/

Label Format

Annotations follow the standard YOLO format (<class> <x_center> <y_center> <width> <height>), normalized between 0 and 1.

• Quality Assurance: All labels were manually annotated and reviewed in Roboflow to ensure tight bounding boxes.

Classes

subset	Class Name	Description
CARS-TRUCKS	pickup	Light vehicles (Camionetas, Service Trucks, Hilux/L200).
MACHINERY	heavy_machinery	Mining equipment (Bulldozers, Excavators, Dump Trucks).
PIEZOMETER	piezometer	Critical monitoring instruments for tailings dam stability.

Preprocessing & Augmentation

To ensure model robustness against lighting changes and drone orientation, the following augmentation pipeline was applied via Roboflow (generating 3 outputs per training example):

• Flip: Horizontal, Vertical.
• Rotation: 90° Clockwise, 90° Counter-Clockwise, Upside Down.
• Color Space: * Saturation: ±25%
  • Brightness: ±10%

Benchmark Results

The dataset has been tested with YOLOv8 architectures using warmup and fine-tuning strategies. The metrics below reflect the performance on the vehicle detection tasks.

Model Architecture	mAP 50-95	mAP 50	Precision	Recall
YOLOv8 Medium (Augmented)	0.773	0.994	0.972	0.989
YOLOv8 XLarge (Augmented)	0.720	0.970	0.997	0.955
YOLOv8 Medium (Base)	0.671	0.990	0.940	0.972
YOLOv8 Nano (1024px)	0.660	0.992	0.993	0.944

Credits Author: Tito Ruiz Haros

Organization: Linkapsis

Context: Operational Topography & GeoAI Development.