dev-aircraft-ident

Practice to build/play with a model to detect aircraft in satellite imagery

• Dataset (building in progress) - https://app.roboflow.com/ufo-optix/ufo-bounding-box-detection

About

This uses the YOLOv8 model due to the following reasons:

1. It's extremely fast.
2. It's very flexible (Formats that PyTorch, Tensorflow, etc. can consume).
3. It's stupid easy to add new weights and classes to existing models and/or create a new model.

Models

All models are located in the models directory.

• August 8th, 2023 Training:
  • 

Dataset Building

In the repository, I am aggregating EO Browser (Located in the raw_images directory)images that contain known aircraft movement defined by an RGB signature on the image. These images are then uploaded to our UFO Bounding Box Detection Dataset where we will then annotate the images and prepare them to be trained against. We will need ~2,000 images for a single class according to darknet.

for each object which you want to detect - there must be at least 1 similar object in the Training dataset with about the same: shape, side of object, relative size, angle of rotation, tilt, illumination. So desirable that your training dataset include images with objects at different: scales, rotations, lightings, from different sides, on different backgrounds - you should preferably have 2000 different images for each class or more, and you should train 2000*classes iterations or more

Image Naming - Each image is named by the ICAO code for the airport/base.

To get access to the training dataset with the ability to add and annotate images, please send us your email in Discord.

Training

Here have been my steps that you can use to reproduce training.

1. Find images of content that has the object you're looking for. You're going to need a-lot of them (100-500 images).
2. Head over to https://app.roboflow.com/ufo-optix/ and create a new project and add your images or add images to an existing project such as the https://app.roboflow.com/ufo-optix/ufo-bounding-box-detection.
3. Annotate your images and save the annotations.
4. Export your Generated Dataset into YOLOv8 format as a ZIP file.
   1. 
5. Unzip and move:
   1. train folder to <path_to_git_repo>/data/train
   2. test folder to <path_to_git_repo>/data/test
6. Make sure the config.yaml has your systems information and any classes (the items you annotated).

   1. path: <PATH_TO_GIT_REPO>\dev-aircraft-ident
      train: data\images\train
      val: data\images\train
      
      nc: 1
      
      names: ['Aircraft']

7. Open the aircraft_detection notebook and run it.
8. Once you've trained your model, you will see a runs directory with the training information and your test cases as images that have been analyzed. This folder will also have your new weights.

Example Datasets

1. For Satellite Aircraft Training, I built my own and used the following:
   1. https://universe.roboflow.com/gdit/aerial-airport
   2. https://app.roboflow.com/joshua-faust-6mncn/satellite-aircraft-detections