harsh-awasthi/bluemockdrone

Blue Mock Drone Frame Object Detection Dataset

Synthetic and auto-annotated bounding box detection dataset for tracking a 3D-printed blue mock drone frame / gimbal wheel in varied real-world environments.

Designed to overcome color-filtering / HSV tracking failures under difficult lighting conditions, heavy shadows, and adversarial blue backgrounds.

---

Dataset Summary

• Total Images: 400 synthetic images (640×640 resolution)
• Train Split: 320 images (images/train/)
• Validation Split: 80 images (images/val/)
• Annotation Format: YOLO normalized bounding boxes (labels/train/, labels/val/)
• Classes:
  • 0: blue_object (3D-printed blue mock drone frame)

---

Dataset Structure

.
├── data.yaml                  # YOLO dataset configuration
├── README.md                  # Dataset Card
├── images/
│   ├── train/                 # 320 synthetic training images
│   └── val/                   # 80 synthetic validation images
├── labels/
│   ├── train/                 # 320 YOLO normalized label txt files
│   └── val/                   # 80 YOLO normalized label txt files
├── cutouts/                   # 5 transparent PNG cutouts across varied angles
├── previews/                  # 16 visual preview samples with bounding boxes
└── scripts/
    ├── extract_cutouts.py     # Script to extract transparent PNG cutouts
    ├── generate_dataset.py    # Script to synthesize and annotate dataset
    └── train_yolo.py          # YOLOv11n one-click training script

---

Augmentations & Synthesis Methodology

1. Background Diversity: Blended over real-world natural, indoor, workshop, and outdoor scenes from the COCO dataset, plus textured wooden desks, gradient surfaces, and adversarial blue backgrounds.
2. Physical Transformations:
   • $360^\circ$ continuous random in-plane rotation.
   • 3D perspective warp and shear ($\pm 25^\circ$).
   • Dynamic scale variations (0.18× to 0.70× relative to frame).
3. Photometric Perturbations:
   • Exposure, contrast, and brightness shifts.
   • Non-uniform directional shadow gradients.
   • HSV color temperature jitter.
4. Sensor & Realism Degradation:
   • Simulated Gaussian blur and directional motion blur.
   • Synthetic CMOS camera ISO sensor noise.
   • JPEG compression artifacts.
5. Adversarial Distractors & Hard Negatives:
   • Blue patches, background shapes, and negative samples (background only without target) to force the detector to learn the distinctive circular 4-spoke geometry rather than relying purely on blue color.

---

How to Train YOLOv11 on this Dataset

1. Install Ultralytics

pip install ultralytics

2. Train using the YOLO CLI

yolo detect train data=data.yaml model=yolo11n.pt epochs=50 imgsz=640 batch=16

3. Train using Python

from ultralytics import YOLO

# Load base YOLOv11 nano detection model
model = YOLO("yolo11n.pt")

# Train on this dataset
results = model.train(
    data="data.yaml",
    epochs=50,
    imgsz=640,
    batch=16
)

---

License

MIT License.