YOLOv8n -- SAR Vessel Detection
YOLOv8 nano model trained for vessel detection in Synthetic Aperture Radar (SAR) imagery.
Quantized and exported for embedded edge deployment (Jetson, Google Coral).
Model Details
- Base model: YOLOv8n
- Task: Object Detection (vessel/ship)
- Dataset: SAR-Ship (Roboflow)
- Input size: 640x640
- Target hardware: Jetson Orin Nano, Google Coral Edge TPU, RTX GPUs
Performance
| Model |
mAP50 |
mAP50-95 |
Precision |
Recall |
Size (MB) |
| PyTorch FP32 |
0.9168 |
0.6646 |
0.9183 |
0.84 |
6.26 |
| TorchScript FP16 |
0.9124 |
0.6271 |
0.9278 |
0.8274 |
12.43 |
Repository Structure
unquantized/
best.pt (PyTorch FP32 - original trained weights)
quantized/
best.onnx (ONNX FP32 - cross-platform inference)
best.torchscript (TorchScript FP16 - Jetson / RTX GPUs)
best_int8.tflite (TFLite INT8 - Google Coral / MCU, Linux only)
evaluation_results.json
evaluation_results.csv
Available Formats
| File |
Format |
Use case |
unquantized/best.pt |
PyTorch FP32 |
Training / fine-tuning / full accuracy inference |
quantized/best.onnx |
ONNX FP32 |
Cross-platform CPU/GPU inference |
quantized/best.torchscript |
TorchScript FP16 |
Jetson Orin / RTX GPU deployment |
quantized/best_int8.tflite |
TFLite INT8 |
Google Coral Edge TPU / microcontrollers |
Usage
PyTorch (unquantized)
from ultralytics import YOLO
model = YOLO("unquantized/best.pt")
results = model("your_sar_image.png")
results[0].show()
ONNX (quantized)
from ultralytics import YOLO
model = YOLO("quantized/best.onnx")
results = model("your_sar_image.png")
results[0].show()
TorchScript FP16 (quantized, RTX / Jetson)
from ultralytics import YOLO
model = YOLO("quantized/best.torchscript")
results = model("your_sar_image.png")
results[0].show()
Training Details
- Optimizer: AdamW
- Epochs: 100 (early stopping, patience=20)
- Augmentation: mosaic, rotation (+-15 degrees), flip -- color aug disabled (SAR is grayscale)
- Quantization: Post-Training Quantization (PTQ) via TorchScript FP16 and TFLite INT8
