☀️ High-Precision Rooftop Detection in Urban Aerial Imagery

This repository implements a hybrid detection–segmentation pipeline for high-precision rooftop detection from aerial imagery. It integrates state-of-the-art object detection models (YOLOv10/11/12, DETR) with the Segment Anything Model (SAM) to guide segmentation. Additionally, baseline semantic and instance segmentation models (U-Net, DeepLabV3+, Mask R-CNN) are benchmarked for comparison.

---

📌 Project Highlights

• 🔄 Two-stage hybrid pipeline: Detection followed by region-guided segmentation.
• 🎯 Prompt-based segmentation: Bounding boxes from detectors used as SAM prompts.
• 🇮🇳 Indian rooftop datasets: Custom-labeled Gandhinagar dataset (600+ images) and upcoming dense-region Ahmedabad dataset.
• 📊 Comprehensive comparison: Benchmarking across rooftops, roads, trees, and solar panels using 4 public datasets.

---

🏗️ Pipeline Architecture

🔹 Baseline Pipeline

Standalone segmentation models trained on aerial imagery:

• U-Net
• DeepLabV3+
• Mask R-CNN
• Segment Anything Model (SAM) (zero-shot and fine-tuned)

📁 Directory: BaseLineModels/

---

🔹 Hybrid Pipeline

Proposed two-stage approach:

1. Detection: YOLOv10/11/12 or DETR localizes rooftops.
2. Prompted Segmentation: SAM generates masks using bounding box prompts.

📁 Directory: Solar_Rooftop_Detection/

---

🌍 Dataset Coverage

• Inria Aerial Image Labeling Dataset (Rooftops)
• Zenodo Solar Panel Dataset
• Massachusetts Roads Dataset
• ISPRS Potsdam Dataset (Trees)

🇮🇳 India-Specific Contribution

• ✅ Gandhinagar Dataset: 600+ manually annotated rooftop images.
• ✅ All models trained and tested on this dataset.
• 🔜 Ahmedabad Dataset: Collected dense urban images—results coming soon.

📁 Directory: Solar_rooftop_Detection_indian_images/

---

📊 Model Performance (Rooftops Example)

Model	Pixel IoU	Dice Score	Region Accuracy
YOLOv12 + SAM	0.780	0.863	0.965
DETR + SAM	0.745	0.840	0.953
U-Net	0.633	0.774	1.000
DeepLabV3	0.648	0.747	0.911
Mask R-CNN	0.017	0.019	0.021

📁 Full results: accuracy.txt, RooftopAccuracy.txt

---

📄 Reports

• 📘 Report_Version_1.pdf: Methodology, architecture, early experiments, 4 urban segmentation tasks, detailed benchmarks, and pipeline diagrams

---

✅ Key Contributions

• 🧱 Designed & implemented a modular rooftop segmentation pipeline
• ✍️ Annotated 600+ Indian rooftops (Gandhinagar)
• 📦 Integrated 4 detectors (YOLOv10–12, DETR) and 4 segmenters (U-Net, DeepLabV3, Mask R-CNN, SAM)
• 🧠 Leveraged SAM for zero-shot and prompt-based segmentation
• 🧪 Conducted detailed evaluation across four public datasets and Indian rooftops

---

Stay tuned for results on the Ahmedabad dataset, focusing on dense urban morphology.

Developed by Dhruv Panchal And Ritwik Agrawal. For questions or collaboration, reach out via GitHub.

For setup, usage, training, and visualization instructions — refer to the upcoming usage section or contact the authors.

---