“The model is only as good as the data.” - Daniel Puratich?

As the AEAC Student UAS Competition CONOPS is different every year, a new model needs to be trained every year as well. Unfortunately, this means a lot of manual labour.

Data Collection

Software

Software should have the same API as the one used in competition (e.g. OpenCV Python)

R D TODO: Publish the software

Video

If video is collected instead, ffmpeg can be used to transform the video file into individual frames:

• ffmpeg -i [input file] frames/%05d.png

  • Requires a directory called frames/ to exist in the same directory as the input file

Hardware

• Flight computer: A Raspberry Pi with a 32GB microSD card is sufficient

• Camera: The same one used at competition (e.g. the $200 CV Camera)

Flight

The drone typically flies at a height of 60 metres above ground level.

Data Cleaning

Once images have been collected, any image with features of interest (e.g. landing pads) are kept, and the rest are discarded. This can be parallelized by uploading the images to a folder on OneDrive and creating a signup sheet so that members process a few hundred images at a time.

• Members can use the Tiles view to see the images without having to click through each one individually

• Images can be deleted in the OneDrive webpage

2022-2023 instructions: Landing Pad Data Cleaning Instructions

Data Labelling

Once images have been cleaned, they are downloaded and then uploaded to Roboflow for labelling. A Roboflow account with a non-existent email can be created (e.g. warg-0a@warg.warg ). Each Roboflow project can only have up to 3 collaborators, so multiple projects can be created. It is unclear whether simultaneous logins cause accounts to be banned, so it is better to create new accounts for every project and to add a signup sheet for members currently logged into the account.

New Roboflow project:

1. Create a Roboflow account (e.g. warg-0a@warg.warg )

   1. https://app.roboflow.com

2. In the main account, create a new workspace (e.g. warglandingpad)

3. In the workspace, create a new project (e.g. warglandingpad)

   1. Typically the use case requires Object Detection

   2. Add the first class; additional classes are added in further instructions

4. Log out

5. Create 2 additional Roboflow accounts (e.g. warg-0b@warg.warg , warg-0c@warg.warg )

6. Log out

7. In the main account, navigate to the workspace (left bar)

8. Invite the additional Roboflow accounts (top right)

9. Log out

10. In the additional Roboflow accounts, accept the invitation (bell icon in top right)

11. Log out

12. In the main account, navigate to the workspace (left bar)

13. Navigate to the project (bottom centre)

14. Uploading:

    1. Upload the images (left bar)

    2. The images will cache locally in the browser (i.e. RAM), so split the images (use a good batch name)

    3. Save and Continue

    4. Repeat until the images for the project are uploaded

15. Classes:

    1. Navigate to Annotate (left bar)

    2. Assign a batch to the main account

    3. Click on an image

    4. Use the bounding box tool (right bar) to draw a box anywhere

    5. Add the new class

    6. Repeat d-e until all classes have been added

    7. Click on the project name (top left)

    8. Click on the 3 dots (left bar) and navigate to Project Settings

    9. Check Lock Annotation Classes

    10. Return to the image that was just labelled

    11. Delete the boxes

Repeat for as many projects are required.

2022-2023 instructions: Landing Pad Data Labelling Instructions

When a batch is labelled, verify the labels are correctly placed by sampling a few images (scroll down and randomly pick an image, repeat). Once the batch is verified, add that batch to the dataset with the default 70/20/10 split.

Data Augmentation

Once all images have been labelled, navigate to Dataset in the left bar and ensure that the split is as close to 70/20/10 as possible (a few images of variance is fine). Then, navigate to Generate.

Generate augmented version:

1. Leave default

2. Leave default

3. None

4. Flip, Hue, Saturation, Brightness.

   1. 2022-2023 used:

      1. Flip: Horizontal, Vertical

      2. Hue: -27° and +27°

      3. Saturation: -75% and +75%

      4. Brightness: -25% and +0%

5. 3x (but if higher multiplier is available, use it)

6. Click Generate. It takes a few minutes for the version to be generated.

Download version:

1. Click on the version to download under Versions

2. Click on Export Dataset (left)

3. Export:

   1. Format: YOLOv8

   2. Select download zip to computer

   3. Uncheck Also train a model etc.

   4. Click Continue

4. Download the zip file

Data Recombination

Extract all zip files to the same location. It is fine to have the README and .yaml files overwrite each other, as the READMEs can be ignored and the .yaml are almost identical (the differences are in the Roboflow section).

Ideally, the image order is randomized. However, labels are required to correspond to the same image.

Roboflow provides hashes at the end of the file names, which is sufficient and can be used as the prefix. Bulk Rename Utility is used to rename files: https://www.bulkrenameutility.co.uk/

Bulk Rename Utility settings:

• Move/Copy Parts (6): Copy last n, 32, To start, 1, Sep.: . (dot)

Apply the rename to all files (test, train, valid).

The 3 directories now contain the dataset and are ready for training.