
Task	Object detection (detect + identify class)
Base model	YOLO26n (COCO pretrained)
Fine-tuned models	3 specialized checkpoints
Hardware used	NVIDIA GeForce RTX 3090 (24 GB)
Framework	Ultralytics / PyTorch

Training images are not shipped in this repo. Only weights, configs, metrics, and a few sample images are included.

Models

Checkpoint	Domain	Classes	Val mAP50	Val mAP50-95	Precision	Recall
`weights/yolo26n_mvrsd_best.pt`	Remote-sensing vehicles (MVRSD)	5	0.842	0.586	0.812	0.792
`weights/yolo26n_vehicles8_best.pt`	Ground vehicles	8	0.644	0.419	0.751	0.553
`weights/yolo26n_aircraft_best.pt`	Military aircraft	74	0.706	0.625	0.770	0.605

Baseline vs fine-tune (same val split)

Model	Baseline mAP50 (COCO YOLO26n)	Fine-tuned mAP50	Gain
MVRSD	0.005	0.842	+83.7 pp
Vehicles8	0.063	0.644	+58.1 pp
Aircraft	0.001	0.706	+70.5 pp

Training recipe (all three): 100 epochs · imgsz=640 · batch=16 · cos_lr · optimizer=auto · start from yolo26n.pt.

Classes

MVRSD — 5 vehicle types

ID	Code	Description
0	SMV	Small Military Vehicle
1	LMV	Light / Large Military Vehicle
2	AFV	Armored Fighting Vehicle
3	CV	Combat / Civilian Vehicle
4	MCV	Medium / Military Cargo Vehicle

Vehicles8 — 8 ground types

APC · ENG · IFV · IMV · MRL · ART · Tanks · Trucks

Aircraft — 74 types

Full list in configs/aircraft.yaml (F16, F35, Su57, AH64, C130, …).

Quick start

Install

pip install ultralytics huggingface_hub

Inference (download from Hub)

python

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

repo = "RehmanKhalid/yolo26-military-detection"

ckpt = hf_hub_download(repo_id=repo, filename="weights/yolo26n_mvrsd_best.pt")
model = YOLO(ckpt)

results = model.predict("your_image.jpg", conf=0.25, imgsz=640)
results[0].show()      # draw boxes
results[0].save("out.jpg")

Switch models

python

# Ground vehicles (8-class)
ckpt = hf_hub_download(repo_id=repo, filename="weights/yolo26n_vehicles8_best.pt")

# Aircraft (74-class)
ckpt = hf_hub_download(repo_id=repo, filename="weights/yolo26n_aircraft_best.pt")

CLI

bash

yolo detect predict \
  model=weights/yolo26n_mvrsd_best.pt \
  source=your_image.jpg \
  conf=0.25 imgsz=640

Repository layout

text

weights/
  yolo26n_mvrsd_best.pt       # 5-class remote-sensing vehicles
  yolo26n_vehicles8_best.pt   # 8-class ground vehicles
  yolo26n_aircraft_best.pt    # 74-class aircraft
configs/
  mvrsd.yaml
  vehicles8.yaml
  aircraft.yaml
  categories10.yaml           # folder labels only (no boxes)
metrics/
  military_id_*.json          # baseline & fine-tuned val reports
examples/
  *_sample.*                  # illustrative images (not the full dataset)
REPORT.md                     # technical write-up (MVRSD experiment)

How evaluation was done

Standard benchmarking (before/after), not score-chasing:

Evaluate COCO-pretrained yolo26n on each dataset val split
Fine-tune on that dataset’s train split
Re-evaluate the same val split with best.pt
Save JSON reports under metrics/

Primary metrics: mAP50, mAP50-95, precision, recall.

Intended use

Research, education, and demos of domain adaptation for detectors
Starting point for further fine-tuning on your own labeled imagery

Not intended for: production targeting, weapons control, or unlawful surveillance.

Limitations

Domain shift: performance drops on sensors / altitudes / countries not seen in training
Small / occluded / partial vehicles are harder (loose boxes more likely)
Nano model (n) favors speed over max accuracy — try larger YOLO26 scales for tighter boxes
Dataset licenses remain with the original dataset owners; this repo redistributes weights only

Citation

Please cite Ultralytics YOLO and the original dataset authors if you use these weights.

bibtex

@software{ultralytics_yolo,
  title = {Ultralytics YOLO},
  author = {Ultralytics},
  url = {https://github.com/ultralytics/ultralytics}
}

License

Model weights and configs in this repository: provided for research/education.
Underlying training datasets: follow their original licenses and terms.
Base architecture: Ultralytics YOLO (see Ultralytics license).