Defense-oriented object detection fine-tune of YOLO26x for detecting people, civilian vehicles, and military assets (ground vehicles, armor, artillery, aircraft, helicopters) in imagery.
Intended for research, simulation, and authorized defense / security analytics only. Not for unlawful targeting or surveillance.
| Architecture | YOLO26x (Ultralytics) |
| Task | Object detection |
| Input | RGB images (default imgsz 640) |
| Framework | Ultralytics / PyTorch |
| Weights | OKIYOLO26X-DEFENSIVE.pt |
person · car · bus · truck · tank · apc · artillery · military_vehicle · aircraft · helicopter
Same validation split and protocol for both checkpoints.
| Metric | YOLO26x (COCO pretrained) | OKIYOLO26X-DEFENSIVE | Δ |
|---|---|---|---|
| mAP50 | 48.0% | 75.8% | +27.8 |
| mAP50-95 | 39.3% | 59.8% | +20.5 |
| Precision | — | 76.4% | — |
| Recall | — | 72.1% | — |
pip install ultralytics huggingface_hub
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
weights = hf_hub_download(
repo_id="theokiGuyOKI/OKIYOLO26X-DEFENSIVE",
filename="OKIYOLO26X-DEFENSIVE.pt",
)
model = YOLO(weights)
results = model.predict("your_image.jpg", imgsz=640, conf=0.25)
results[0].show()
CLI:
yolo predict model=OKIYOLO26X-DEFENSIVE.pt source=your_image.jpg imgsz=640
Fine-tuned on an aggregated mix of publicly available object-detection datasets (aerial / ground military & civil vehicles). Sources include:
| Source | Notes |
|---|---|
| Military Vehicles Detection (Roboflow Universe) | CC BY 4.0 |
| Military Tanks (Roboflow Universe) | CC BY 4.0 — tanks & aircraft types |
| Military-Civil-Vehicles (Roboflow Universe) | CC BY 4.0 |
| Russian-military-annotated (Roboflow Universe) | CC BY 4.0 — credit: Tuomo Hiippala / Univ. of Helsinki |
| Army_vehicle_yolov8 (Roboflow Universe) | MIT |
| MVRSD (Military Vehicle Recognition / Surveillance Dataset) | Public military-vehicle imagery + labels |
Labels were remapped into the 10 unified classes above. This repository distributes model weights only — underlying images/labels are not re-uploaded here; obtain them from the original publishers under their licenses.
AGPL-3.0 (inherits Ultralytics YOLO licensing for this stack). Commercial use may require an Ultralytics Enterprise license — check Ultralytics licensing.
@misc{okiyolo26x-defensive,
title = {OKIYOLO26X-DEFENSIVE},
author = {theokiGuyOKI},
year = {2026},
url = {https://huggingface.co/theokiGuyOKI/OKIYOLO26X-DEFENSIVE}
}