Skip to content

Latest commit

 

History

History
85 lines (62 loc) · 3.97 KB

README.md

File metadata and controls

85 lines (62 loc) · 3.97 KB
Raw

data-labelling

Setup

1. Clone repository

Run the command

git clone https://github.com/nsw-wildlife-drone-hub/data-labelling.git

Or alternatively manually download the code by clicking Code -> Download ZIP. The downloaded folder will then need to be unzipped and saved in a workspace.

Labelling

1. Download Data

  1. Determine a suitable detection to label in the Detection Labelling List and assign your name in the Labeller column
  2. Download the .MP4 video and any jpg screenshot available
  3. Move the downloads into the Data/ folder in the data-labelling workspace
  4. Create a new folder by the same name as the labelling_foldername column in the Detection Labelling List
  5. Also rename the .MP4 video basename to the labelling_foldername

2. DarkLabel

First use the available screenshot to determine the target detection to label. the time_in_video column in the Detection Labelling List is also a useful guide to how far in the video is the detection visible.

To begin creating labels:

  1. Open DarkLabel
  2. Click 'Open Video' and select the .MP4 video
  3. Maximize the window panel (this will give finer control of the mouse)
  4. Scroll to when the detection first appears
  5. Create a label
  • Left-click: draw a box
  • Shift+Right-click: delete box
  • Shift+Left-click: adjust box
  • Ctrl+Scroll: zoom in/out
  • Enter: predict next box
  1. Save your Ground-truths (GT) by clicking 'GT Save As' and save in the new folder that was created

2.1 Good vs bad labels

  1. Make tight boxes: make the borders of the box touch the object, don't leave white space and definitely don't crop the object
  2. Do not abuse the predict tool: it can save a lot of time but be sure to 'reset' the label manually after a number of frames so inaccuracies do not compound
  3. Label occluded objects: if the object is partially hidden, ask yourself, can I still identify the animal in the picture?
  4. Take your time: the accuracy of your label will affect the accuracy of the AI. Take a break if you need it, it will be worth it for everyone.

3. Extracting Images and Videos

Double-click dataextect.exe. This will automatically extract images and videos from the data. There will be 3 scans run and printed:

  1. Folders containing labels: these are folders that have labels (.txt) found in them
  2. Folders with no images: these are folders with labels (.txt) but no images (.jpg)
  3. Matching video data: these are videos with the same name as the folder

Each time a labelling process is completed, there will be a dataset found to extract, otherwise, the program will ask to terminate. If no dataset is found to be extracted, check inside the label folder and also check the name of the folder and the video.

4. Compress labels

(Option A)

  1. Zip the folder using 7-zip

(Option B)

  1. Right-click the labels+images folder

  2. Select Send to -> Compressed (zipped) Folder

5. Upload data

  1. Upload the compressed folder to the Data/ folder in Gdrive

  2. Upload the output_gt.mp4 data to the Review_AI folder in Gdrive

  3. Upload the output.mp4 data to the Review_Pilot folder in Gdrive

Monitoring and Reviewing

Labelling Pipeline

On the Detection Labelling List, there will be a review procedure made by a pilot and the ML engineer to approve the quality of the data that is labelled. If at any stage, the quality does not pass the reviews, a note needs to be made by the reviewer as to why it was rejected. Following the review, the detection will be relabelled by the same labeller. The labels can be simply modified by GT Load into the DarkLabel and adjusting labels according to the notes.

Packaging .py into .exe

Run the following commands to convert the python file into a packaged executable.

python -m venv pyinstall
.\pyinstall\Scripts\activate
pip install -r requirements.txt
pyinstaller --onefile dataextract.py
mv dist\dataextract.exe dataextract.exe -force