Blue Mock Drone YOLOv11n Object Detector

High-precision, ultra-fast YOLOv11 Nano object detection model trained to detect and track a 3D-printed blue mock drone frame / gimbal wheel across varied physical environments, complex backgrounds, and challenging lighting conditions.

Dataset: harsh-awasthi/bluemockdrone

Performance Summary (Epoch 50)

Evaluated on 80 validation images ($640 \times 640$ resolution):

Metric	Score	Note
mAP@50	99.3% (`0.993`)	Outstanding detection precision
mAP@50-95	95.6% (`0.956`)	Tight multi-threshold bounding box alignment
Precision	96.8% (`0.968`)	Near-zero false positive rate
Recall	98.0% (`0.980`)	Detects virtually all object instances
Inference Latency	7.4 ms / frame	Over 135 FPS on Apple Silicon (M5)
Parameters	2,582,347 (2.58M)	Ultralight footprint for embedded edge deployment
FLOPs	6.4 GFLOPs	Real-time capable on Raspberry Pi & edge AI

Training History & Loss Curves

Model Formats Included

best.pt: Native PyTorch / Ultralytics checkpoint (5.5 MB).
best.onnx: Slimmed ONNX format (opset 18, 10.1 MB) for OpenCV DNN, ONNX Runtime, TensorRT, and ROS integration.

Quickstart & Usage

1. Using Ultralytics Python API

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# Download weights from Hugging Face
model_path = hf_hub_download(repo_id="harsh-awasthi/bluemockdrone", filename="best.pt")

# Load model
model = YOLO(model_path)

# Predict on an image, video, or webcam
results = model.predict(source="image.jpg", conf=0.5, save=True)
for r in results:
    print(r.boxes.xyxy)  # Bounding boxes
    print(r.boxes.conf)  # Confidence scores

2. Using OpenCV DNN with ONNX

import cv2
from huggingface_hub import hf_hub_download

# Download ONNX weights
onnx_path = hf_hub_download(repo_id="harsh-awasthi/bluemockdrone", filename="best.onnx")

# Load via OpenCV DNN
net = cv2.dnn.readNetFromONNX(onnx_path)
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_OPENCV)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_CPU)

3. Using the Ultralytics CLI

# Direct prediction
yolo predict model=best.pt source=0 conf=0.5

Training Details

Base Architecture: yolo11n.pt
Input Resolution: 640 × 640
Epochs: 50
Batch Size: 16
Hardware: Apple Silicon GPU via Metal Performance Shaders (mps)
Training Time: ~14 minutes (0.23 hours)

License

MIT License.

Downloads last month: -

harsh-awasthi
/

bluemockdrone