Modelpublic

AIOmarRehan/sived_yolov8m_obb_best

sourceHugging Facemitupdated 6mo agoView on Hugging Face

1likes

Model Card

Model Card: SIVED YOLOv8m-OBB

Model Details

Property	Value
Architecture	YOLOv8m-OBB (Oriented Bounding Box)
Parameters	~25.9M
Base Weights	YOLOv8m-OBB pretrained on DOTAv1
Fine-Tuned On	SIVED (SAR Image Vehicle Detection)
Task	Oriented object detection
Classes	1 (Vehicle)
Input Size	640 x 640 px
Framework	Ultralytics 8.4.37, PyTorch 2.10.0
File	`sived_yolov8m_obb_best.pt` (~53 MB)

Intended Use

Detection of vehicles in Synthetic Aperture Radar (SAR) imagery using oriented bounding boxes. Designed for remote sensing applications including surveillance, defense, disaster response, and urban monitoring.

Training

Parameter	Value
Optimizer	AdamW
Learning Rate	0.001 (cosine decay to 0.01x)
Batch Size	16
Epochs	100 max (early stop at 70, best at ~50)
Patience	20
Mosaic	1.0
Copy-Paste	0.2
Rotation	15 degrees
HSV Hue / Saturation	0.0 / 0.0 (disabled for grayscale SAR)
HSV Value	0.4
Hardware	Tesla T4 (16GB VRAM), Google Colab

Performance (Test Set, 103 images)

Metric	Value
mAP@0.5	0.988
mAP@0.5:0.95	0.816
Precision	0.970
Recall	0.979
F1 (peak)	0.97 at conf=0.494

Baseline Comparison

The pretrained DOTAv1 model scored 0.0 on all metrics when evaluated on SIVED test data. DOTAv1 contains optical aerial categories with no SAR vehicle mapping, confirming that domain-specific fine-tuning is essential.

Limitations

—Trained on a single class (Vehicle) only.
—Optimized for 512 x 512 grayscale SAR imagery. Performance on optical images or different resolutions has not been evaluated.
—The SIVED dataset contains 1,044 images from three radar sources (FARAD, MiniSAR, MSTAR). Generalization to other SAR sensors or bands is not guaranteed.

Citation

If you use this model, please cite the original SIVED dataset:

bibtex

@Article{rs15112825,
  author  = {Lin, Xin and Zhang, Bo and Wu, Fan and Wang, Chao and Yang, Yali and Chen, Huiqin},
  title   = {SIVED: A SAR Image Dataset for Vehicle Detection Based on Rotatable Bounding Box},
  journal = {Remote Sensing},
  volume  = {15},
  number  = {11},
  pages   = {2825},
  year    = {2023},
  doi     = {10.3390/rs15112825}
}