annotationSteps.md

Workshop Annotation Steps

Overview

1. Background Information
2. Create Annotation Project
3. Perform Annotations
4. Save Annotations

Background information

CVAT projects have a hierarchy incorporating tasks and jobs. Annotation projects may have many annotation tasks and jobs. Tasks are where you upload video footage and specify the associated project and labels. Tasks represent the progress for annotation footage and details of the task. They also allow you to upload existing annotations and export annotations. Tasks may have multiple jobs. Jobs allow you to split up tasks by frame to divvy up the annotations among a team.

Steps to follow to create an annotation project

1. Create an account and login at cvat.org.

While creating a new account, a valid email address is required. Tool requires email confirmation by clicking on the link sent to this email address

2. Use the top navigation bar to click Projects.

3. Click the blue + sign and then Create a new project.

4. Copy the details from the image below to fill out the project to be created and select Submit & Open. For the raw label data, copy/paste the data snipped below the picture. These data represent the labels & their attributes.

   

   Raw Label data to copy/paste into the CVAT UI:

[
  {
    "name": "Person",
    "id": 1199059,
    "color": "#c06060",
    "attributes": [
      {
        "id": 1198651,
        "name": "Identity",
        "input_type": "number",
        "mutable": false,
        "values": [
          "1",
          "1000",
          "1"
        ]
      },
      {
        "id": 1198650,
        "name": "Occlusion",
        "input_type": "number",
        "mutable": false,
        "values": [
          "0",
          "2",
          "1"
        ]
      }
    ]
  }
]

Note: For more information on the labels used for the workshop, please refer to the Annotation Specifications Doc.

5. Click the blue + sign at the bottom right of the newly created project to create a new task for it.
6. Create your task copying the details below and click Submit & Open on your task:

   Note: The test footage may be found here. You will use this video file for your annotation sample data. It may take a few seconds-minutes to upload the video footage pending the network bandwidth.

7. Under the job details, assign the job to your username as seen below.
8. Click on the job # that was created above. This will take you to a screen where you can start data annotations. This is where you will use the supporting documents to perform your data annotations. Feel free to ask questions as needed during this time, and collaborate with those around you as you progress.

Steps to follow to perform data annotations

1. Hover over the rectangle shape on the left navigation bar, and select Track.

Note: We are annotating a video file. This means we will want to track people as they move across the frame throughout the video.

2. Place your cursor at the top left of the person and click. This should create the start of your bounding box. Now go to the bottom right of the person to complete your person bounding box, and click when you have included the bottom-rightmost part of your person.
3. In the right most part of the CVAT UI, under the Objects tab, you will leave Identity as 1, and Occlusion as 0 as this is the first person we are creating a bounding box around. This person also has no occlusions impacting the bounding box.
4. Once you have completed annotations for one person, then move onto the next person in the frame.
5. Repeat steps 1-3 for each person going left to right in the frame.

Note: Since Person 2 is seated and we only see ~50% of their body, we will mark their Occlusion field as 1.

Also make sure to increment the number in Identity field for each new person. Refer to the point 3 in Annotation Best Practices for more details on this Identity field

Remember to refer to the Annotation Specifications as needed to adjust your annotations based on occlusions.

When annotations are complete for frame 0, it should look like the below picture.

6. Now that all people in the frame are annotated, click Save to save the current annotations.

7. The track bar (marked in red below) can be used to scroll through the frames one after the other.

   Click on the double right arrow to increment the frame count by 10.

Note: You will see that the initial bounding boxes are no longer properly placed. The bounding boxes size and placement will need to be adjusted since this is a different frame, and people have moved. Also, by annotating every 10 frames, we are leveraging CVAT's interpolation feature. The interpolation feature allows annotators to annotate key frames, and then the frames between them will be interpolated automatically. Each frame annotated is automatically denoted as a key frame with the star shown below.

8. Adjust the bounding boxes for the frame so that they look similar to the picture below.
9. Continue steps 1-7 for every 10 frames of the video annotated each person in the frame.

Steps to follow once frames have been annotated

The following steps are meant to demonstrate to workshop attendees that data annotations go beyond a one-person effort. Data annotation efforts may be reviewed by other team members, and even exported upon completion.

1. Click Tasks.
2. Open the Annotate workshop footage task.
3. Update the Job Stage to validation.

Note: This is an important step when participating in data annotations within a team environment. Updating the stage from annotation to validation signals to others that the current annotations are ready for others to review. Once they have been reviewed and adjusted as necessary, then the annotations may be marked accepted.

4. Click Tasks again.
5. Click on the three dots at the bottom right of the workshop task, and select to Export task dataset.
6. Fill in the export details following the image below.

Note: You will now see the annotations downloaded to a zip folder in the Downloads folder. The zip folder contains an annotations XML formatted file that encodes the annotated data. Different formats may be selected pending the format needed to train the specified model of choice. More information may be found in the section on Annotation Format. The file contains tags corresponding to the project pieces created for this workshop (ie <task>, <labels>, <track>, <box>). Those tags correspond to the task, label, and track ID, bounding box of the people annotated, etc. This exported data is what would be used to train a model.