Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,79 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: mit
|
| 4 |
+
library_name: ultralytics
|
| 5 |
+
tags:
|
| 6 |
+
- military
|
| 7 |
+
- drone
|
| 8 |
+
- object-detection
|
| 9 |
+
- yolo11
|
| 10 |
+
- aerial-surveillance
|
| 11 |
+
- defense
|
| 12 |
+
metrics:
|
| 13 |
+
- map
|
| 14 |
+
- precision
|
| 15 |
+
- recall
|
| 16 |
+
model_name: VeritaMilitary
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
# VeritaMilitary 🛡️🚤✈️
|
| 20 |
+
**VeritaMilitary** is a high-performance, lightweight object detection model specifically optimized for **Aerial Military Surveillance**. It is designed to be deployed on tactical drones with limited onboard computing power.
|
| 21 |
+
|
| 22 |
+
Developed by **M Mashhudur Rahim (XythicK)** at **Arkito Lab** (A Non-Profit Research Organization).
|
| 23 |
+
|
| 24 |
+
## 🚀 Model Overview
|
| 25 |
+
- **Model Architecture:** YOLOv11-Nano
|
| 26 |
+
- **Model Size:** 5.2 MB
|
| 27 |
+
- **Target Domain:** Aerial/Drone View
|
| 28 |
+
- **Inference Speed:** ~2.7ms per image
|
| 29 |
+
|
| 30 |
+
## 📊 Training Results
|
| 31 |
+
The model was trained for **100 epochs**, achieving a solid balance between speed and accuracy for edge devices.
|
| 32 |
+
|
| 33 |
+
| Metric | Value |
|
| 34 |
+
| :--- | :--- |
|
| 35 |
+
| **mAP50** | **40.2%** |
|
| 36 |
+
| **Precision** | **52.6%** |
|
| 37 |
+
| **Recall** | **38.0%** |
|
| 38 |
+
|
| 39 |
+
### Performance Visualization
|
| 40 |
+
The following graphs illustrate the training progress and model evaluation:
|
| 41 |
+
|
| 42 |
+
#### Training Curves
|
| 43 |
+
|
| 44 |
+
**
|
| 45 |
+
|
| 46 |
+
#### Confusion Matrix
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
## 🎯 Domain Specialization (Aerial vs Ground)
|
| 50 |
+
**VeritaMilitary** is highly specialized for top-down perspectives. While it may show sensitivity to perspective shifts in ground-level imagery, it demonstrates exceptional precision in aerial views with confidence scores up to **85%**.
|
| 51 |
+
|
| 52 |
+
| Aerial Detection (Success) | Ground Detection (Limitation) |
|
| 53 |
+
| :---: | :---: |
|
| 54 |
+
|  |  |
|
| 55 |
+
| **85% Confidence in Tank** | **Domain Mismatch (Aerial Optimized)** |
|
| 56 |
+
|
| 57 |
+
## 📦 Multi-Platform Deployment
|
| 58 |
+
The model has been exported to multiple formats to support a wide range of hardware:
|
| 59 |
+
- **ONNX**: Universal CPU/GPU inference.
|
| 60 |
+
- **TFLite**: Mobile and Android Drone controllers.
|
| 61 |
+
- **CoreML**: Apple/iOS devices.
|
| 62 |
+
- **TensorRT**: High-speed NVIDIA Jetson deployment.
|
| 63 |
+
- **OpenVINO**: Optimized for Intel processors.
|
| 64 |
+
|
| 65 |
+
## 🛠 Usage
|
| 66 |
+
```python
|
| 67 |
+
from ultralytics import YOLO
|
| 68 |
+
|
| 69 |
+
# Load VeritaMilitary
|
| 70 |
+
model = YOLO('best.pt')
|
| 71 |
+
|
| 72 |
+
# Run Inference
|
| 73 |
+
results = model.predict(source='drone_footage.mp4', imgsz=640, conf=0.25)
|
| 74 |
+
results[0].show()
|
| 75 |
+
```
|
| 76 |
+
🏢 About Arkito Lab
|
| 77 |
+
Arkito Lab is a non-profit organization dedicated to open-source research and the development of AI solutions for humanitarian and defense technology.
|
| 78 |
+
|
| 79 |
+
Maintained by: M Mashhudur Rahim (XythicK)
|