Night Sky Image Classification & Object Detection

This project provides two machine learning capabilities:

1. Image Classification: Classify entire images into categories (fireballs, meteors, no_fireballs, etc.)
2. Object Detection: Detect and localize objects within images using Faster R-CNN with bounding boxes

🆕 Object Detection (NEW!)

Train a Faster R-CNN model to detect and localize objects in images. Perfect for identifying multiple objects in a single image with precise bounding boxes.

Quick Start:

• See QUICKSTART_OBJECT_DETECTION.md for a 5-minute guide
• See OBJECT_DETECTION_README.md for complete documentation

Features:

• Faster R-CNN with ResNet50 backbone
• Train from scratch on custom datasets
• YOLO format annotation support (via image-classification app)
• Comprehensive metrics: mAP, IoU, Precision, Recall
• Model export to ONNX and TorchScript

---

Image Classification

This project uses a convolutional neural network (CNN) to classify images of the night sky into categories.

Setup

1. Create and activate a virtual environment:

   python3 -m venv venv
   source venv/bin/activate

2. Install dependencies:

   • For common python libs

     pip install -r requirements.txt

   • For systems with a compatible NVIDIA GPU:

     pip install -r torch_requirements.txt

   • For systems with CPU only:

     pip install -r torch_requirements_cpu.txt

Usage

1. Train the model:

   Place your training data in a directory (e.g., training_data) with subdirectories for each class (e.g., fireballs, meteors, no_fireballs, lightnings). Then run the training script:

   python -m src.train training_data
   OR
   python -m src.train_resnet18 training_data

   Set model name in config.py.

       CLASSES = ['fireballs', 'meteors', 'no_fireballs', 'lightnings']
       CLASSES_SIZE = len(CLASSES)
       MODEL_PATH = 'night_sky_model_v11.pth' # used for main and watcher scripts
       TRAIN_MODEL_PATH = 'night_sky_model_v11.pth' # used to saved model file by train.py script
       RESNET18_MODEL_PATH = 'resnet18_trained_v1.pth' # used for main and watcher scripts
       RESNET18_TRAIN_MODEL_PATH = 'resnet18_trained_v1.pth' # used to saved model file by train.py script
       NUM_EPOCHS = 60
       BATCH_SIZE = 8
       LEARNING_RATE = 0.001
   

2. Classify a directory of images (one-time):

   python -m src.main /path/to/your/images

3. Run the automatic watcher service:

   To monitor a folder for new images and classify them automatically, run the watcher script. This will also send Telegram notifications for detected fireballs.

   Before you start: You must edit the src/watcher.py file and replace the placeholder values for TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID with your own credentials.

   python -m src.watcher /path/to/your/source_folder

Building Executables

This project can be packaged into a single executable file using PyInstaller.

1. Install PyInstaller:

   pip install pyinstaller

2. Build the executable:

   The entry point for the executable is run.py. To build the watcher service, you will need to modify run.py to call watcher.main() instead of main.main().

   pyinstaller --name classify_sky_watcher --onefile --add-data "night_sky_model.pth:." run.py

=======

Run docker commands:

Build the image and start services

docker-compose build docker-compose up

Basic data flow diagram:

flowchart 
    A[Cams]-->B[Dropbox]    
    B[DropBox] <-->|files auto sync| C(Local Machine Linux)
    C --> D{Manual Python script running 
    python -m src. ...}
    D --> E[Watcher]
    D --> F[Classify images]
    D --> G[Train model]
    G <--> I@{ shape: docs, label: "Dataset" }
 

Loading