📘 Small Object Detection with Detectron2

---

🔰 Overview

This project focuses on detecting small objects in images using deep learning techniques with Detectron2.
Small object detection remains a challenging task due to scale variation, occlusion, and limited pixel information — yet it’s essential for:

• 🚗 Traffic and surveillance analysis
• 🛰️ Aerial and satellite imaging
• 🩺 Medical image interpretation
• 🤖 Autonomous vehicles and robotics

The notebook walks through all key stages: dataset preparation, model configuration, training, and evaluation.

---

📦 Dataset

This project is dataset-agnostic — any COCO-style annotated dataset with small object instances can be used.

Example dataset structure:

dataset/
├── train/
│   ├── images/
│   └── annotations.json
├── val/
│   ├── images/
│   └── annotations.json

Each annotation file follows the COCO format, including bounding boxes, segmentation masks, and class labels.

---

⚙️ Installation

Clone the repository and install the required dependencies:

git clone https://github.com/your-username/small-object-detection.git
cd small-object-detection
pip install -r requirements.txt

Or install manually:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install detectron2 opencv-python matplotlib tqdm

---

🚀 Usage

To train and evaluate the model, open the notebook:

jupyter notebook small-object-detection.ipynb

Or, if converted to a Python script:

python train_detectron2.py

---

🧹 Data Preprocessing

Before training, the input data is preprocessed to ensure quality and consistency:

✅ Dataset Validation

• Verify image–annotation alignment
• Remove empty or corrupted samples

✅ Resizing and Normalization

• Scale images while maintaining aspect ratio
• Normalize pixel intensity for Detectron2’s input format

✅ Annotation Conversion

• Convert to COCO JSON format
• Verify bounding box and category consistency

---

🧠 Model Configuration & Training

Training is performed using Detectron2, a Facebook AI Research (FAIR) framework.

🧩 Model Selection

• Base architecture: faster_rcnn_R_50_FPN_3x (ResNet-50 backbone with FPN)
• Pretrained weights from COCO used for transfer learning

⚙️ Training Parameters

• Learning rate: 0.00025
• Batch size: 4
• Iterations: ~5000
• Augmentation: random flips and resizing

📏 Evaluation Metrics

• mAP (mean Average Precision) across IoU thresholds
• Separate evaluation for small, medium, and large objects

💾 Output Checkpoints

• Model checkpoints saved in /output after every epoch
• Final model stored as model_final.pth

---

📊 Results & Visualization

After training, performance is analyzed both quantitatively and visually.

📈 Evaluation Metrics

• Example mAP (small objects): ~0.35
• Example mAP (medium/large): ~0.60

🖼️ Detection Visualization

Bounding boxes and labels plotted over sample images.
Model predictions compared to ground truth for validation.

Example output visualization:

from detectron2.utils.visualizer import Visualizer
visualizer = Visualizer(image[:, :, ::-1], metadata=metadata)
out = visualizer.draw_instance_predictions(outputs["instances"].to("cpu"))
plt.imshow(out.get_image()[:, :, ::-1])

---

🧪 Troubleshooting

Common issues and fixes:

❌ CUDA Out of Memory

• Reduce batch size or image resolution

❌ Dataset Registration Error

• Ensure correct paths and annotation file format in DatasetCatalog.register()

❌ Low Detection on Small Objects

• Use Feature Pyramid Network (FPN)
• Apply higher input resolution
• Consider multi-scale training

---

🚀 Deployment Options

You can deploy the trained model using:

✅ Local Inference Script

Run inference on an image locally:

python inference.py --image path/to/image.jpg --model output/model_final.pth

✅ Web or API Deployment

• Use Flask or FastAPI for REST inference
• Convert to ONNX or TorchScript for optimized runtime

---

📦 Deliverables

This repository includes:

• ✅ Jupyter notebook → small-object-detection.ipynb
• ✅ Model checkpoints → /output/model_final.pth
• ✅ Visualization samples → /visuals/
• ✅ Config and training logs → /configs/
• ✅ Requirements file → requirements.txt

---

🏁 Conclusion

✅ Objective Achieved

• Successfully trained a small-object detector using Detectron2

✅ Key Features

• Adaptable to any COCO-style dataset
• Visual evaluation and mAP metrics integrated

✅ Impact

• Enhances object detection for low-visibility and small-scale targets
• Useful for applications in surveillance, remote sensing, and medical imaging

✅ Future Work

• Experiment with custom backbones (e.g., Swin Transformer, ConvNeXt)
• Explore attention-based architectures for improved small object recall
• Optimize inference speed with quantization or pruning

---

⚖️ License

This work is licensed under a Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International License.
Use in CVs, portfolios, or derivative works is not permitted without explicit permission from the author.