Camouflaged Target Detection (CTD)

This project focuses on the detection of camouflaged military vehicles using a combination of YOLOv8 for object detection and CLIP for zero-shot classification. It includes scripts for training a baseline model, fine-tuning on high-resolution images, exporting the model to ONNX, and a Streamlit-based web application for interactive detection.

Features

• Baseline Training: Train a standard YOLOv8 model on your dataset.
• High-Res Fine-Tuning: Fine-tune the model on higher-resolution images (1280px) to improve detection of small, camouflaged objects.
• ONNX Export: Export the trained model to ONNX format with specific opset versioning for broad compatibility.
• Interactive Web App: A Streamlit application that allows users to upload images, run detections, and classify objects using CLIP (Zero-Shot Learning).

Installation

Prerequisites

• Python 3.8 or higher
• CUDA-capable GPU (recommended for training)

Dependencies

1. Clone this repository or navigate to the project directory.
2. Install the required Python packages:

pip install ultralytics streamlit transformers torch torchvision pillow opencv-python-headless numpy

(Note: Ensure you satisfy the specific PyTorch requirements for your CUDA version if you plan to use GPU acceleration.)

Usage

1. Training the Baseline Model

To train the initial YOLOv8m baseline model:

python 01_train_baseline.py

This script uses datasets/military_vehicles/data.yaml and saves results to runs/detect/baseline_run/.

2. Fine-Tuning for High Resolution

To fine-tune the baseline model on larger images (1280px) with memory optimizations (batch size 4):

python 02_finetune_hires.py

This picks up the best weights from the baseline run and saves the fine-tuned model to runs/detect/high_res_finetune/.

3. Exporting to ONNX

To export the fine-tuned model to ONNX format (opset 12):

python 03_export_model.py

4. Running the Web Application

To start the Streamlit interface for testing the model:

streamlit run app.py

This will launch a local web server (usually at http://localhost:8501) where you can upload images and see the detection results.

Project Structure

• 01_train_baseline.py: Script for initial model training.
• 02_finetune_hires.py: Script for high-resolution fine-tuning.
• 03_export_model.py: Utility to export the model to ONNX.
• app.py: The main Streamlit application file.
• datasets/: Directory containing your dataset configuration and images.
• runs/: Directory where training results (weights, logs) are saved.

Credits

• YOLOv8 by Ultralytics
• CLIP by OpenAI