PUSHPENDAR/hrsid-ship-detection

🚢 Ship Detection in SAR Imagery — Faster R-CNN (HRSID)

A Faster R-CNN model trained on the HRSID dataset for robust ship detection in high-resolution Synthetic Aperture Radar (SAR) imagery. Upload a SAR image to detect ships in real time.

---

🧠 Model Architecture

Component	Detail
Framework	Detectron2
Detector	Faster R-CNN
Backbone	ResNet-50 + FPN
Classes	1 (Ship)
Input Size	1400 × 1400
Inference	CPU-compatible (HF Spaces free)

---

📊 Training Configuration

Parameter	Value
Dataset	HRSID (train / val / test)
Train Images	2,914
Val Images	728
Test Images	5,604
Total Instances	16,951 ships
Max Iterations	5,000
Batch Size	2
Base Learning Rate	0.00025
Optimizer	SGD
Eval Period	Every 500 iterations
Checkpoint Period	Every 500 iterations
Score Threshold	0.5
GPU	Tesla T4 (15.6 GB VRAM)

Augmentations: Resize to 1400 × 1400, random horizontal flip.

---

📈 Evaluation Results

Evaluated on the HRSID test split (5,604 images) using COCO-style metrics.

Metric	Value
AP (IoU=0.50:0.95)	47.39
AP50 (IoU=0.50)	71.88
AP75 (IoU=0.75)	56.53
APs (small objects)	48.54
APm (medium objects)	48.55
APl (large objects)	23.17
AR @ maxDets=1	24.30
AR @ maxDets=10	49.20
AR @ maxDets=100	52.40
AR small @ maxDets=100	50.60
AR medium @ maxDets=100	66.70
AR large @ maxDets=100	44.60

---

🛰 Dataset

HRSID (High-Resolution SAR Images Dataset) is a benchmark for ship detection and instance segmentation in SAR imagery.

• 5,604 SAR image crops
• 16,951 annotated ship instances (bounding boxes + segmentation masks)
• Multi-scale vessel distribution across coastal and open-sea scenarios
• Derived from Sentinel-1, TerraSAR-X, and TanDEM-X satellite imagery

---

⚙️ Inference

from detectron2.config import get_cfg
from detectron2.engine import DefaultPredictor
import cv2

cfg = get_cfg()
cfg.merge_from_file("config.yaml")
cfg.MODEL.WEIGHTS = "model_final.pth"
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
cfg.MODEL.DEVICE = "cpu"  # or "cuda"

predictor = DefaultPredictor(cfg)

image = cv2.imread("test_sar.jpg")
outputs = predictor(image)

instances = outputs["instances"]
print(f"Detected {len(instances)} ships")
print(f"Boxes:  {instances.pred_boxes}")
print(f"Scores: {instances.scores}")

---

📦 Repository Files

File	Description
model_final.pth	Trained Faster R-CNN weights
config.yaml	Detectron2 model configuration
labels.json	Class label mapping {0: "ship"}
app.py	Gradio inference app
requirements.txt	Python dependencies

---

🚀 Run Locally

git clone https://huggingface.co/PUSHPENDAR/hrsid-ship-detection
cd hrsid-ship-detection
pip install -r requirements.txt
python app.py

---

📝 Citation

@article{wei2020hrsid,
  title={HRSID: A High-Resolution SAR Images Dataset for Ship Detection and Instance Segmentation},
  author={Wei, Shunjun and Zeng, Xiangfeng and Qu, Qizhe and Wang, Mou and Su, Hao and Shi, Jun},
  journal={IEEE Access},
  volume={8},
  pages={96962--96980},
  year={2020},
  publisher={IEEE}
}

---

📄 License

Apache 2.0 — see LICENSE for details.