Annotating a YOLOv8 Dataset Using Label Studio

Abstract

This article will explain how to annotate a YOLOv8 dataset using Label Studio, as well as how to extract image frames from video files using Python.

Download/Install

• (Required) First, create a Python virtual environment (Python version 3.7+ required) to ensure the project is isolated from other environments. Here, we use the virtualenv included with Pycharm to create it:

Creating a Virtual Environment with PyCharm

• (Required) After creating the virtual environment, open the terminal and enter pip install label-studio opencv-python to install the necessary dependencies.

pip install label-studio opencv-python

Extracting Frames from Video Files

• (Optional) Create and run the segmentation.py file to extract frames from video files.

import cv2


def main(source: str, s: int = 60) -> None:
    """
    :param source:  Video Files
    :param s:       Frame extraction interval, defaulting to saving one frame every 60 frames
    :return:
    """
    video = cv2.VideoCapture(source)
    frame_num = 0
    success, frame = video.read()
    while success:
        if frame_num % s == 0:
            cv2.imwrite(f"./images/{frame_num // s}.png", frame)
        success, frame = video.read()
        frame_num += 1
    video.release()
    cv2.destroyAllWindows()


if __name__ == '__main__':
    main('./videos/sample.mp4')

Image Annotation

• (Required) Enter label-studio start in the terminal to start the service, and then open a browser to access the Label Studio page.

label-studio start

• (Required) First-time users of Label Studio need to register and log in.

To register a Label Studio account

• (Required) After logging into Label Studio, click the "Create" button in the top right corner to create a project.

To create a Label Studio project

• (Required) Enter the project name and description.

Set up basic project information

• (Required) Import the data to be annotated.

Import project annotation data

• (Required) Configure the project labels: Computer Vision -> Object Detection with Bounding Boxes and click the "Save" button in the upper right corner to save.

Configure project labels

• (Required) Click on the unannotated images to begin labeling.

Select data for annotation

• (Required) Press the corresponding number for the label and drag the mouse to annotate. After completing the annotation, click the "Submit" button in the lower right corner to submit.

Annotate data

Export Dataset

• (Required) Click the Export button in the top right corner, select YOLO as the data format to export, and click the Export button again. The browser will then automatically download the annotated dataset.

Export dataset

• (Required) Unzip the downloaded dataset. The directory structure is as follows:

.
│ classes.txt
│ notes.json
│
├─images
│   a4d2cef8-hutao_bg.jpg
│
└─labels
    a4d2cef8-hutao_bg.txt

• (Required) Create three folders named train, test, and val inside the images and labels folders. Distribute the files among these folders in an 8:1:1 ratio. The organized directory structure is as follows:

Note: The filenames in the images and labels subfolders must correspond exactly to each other.

.
│ classes.txt
│ notes.json
│
├─images
│  ├─test
│  │   a4d2cef8-hutao_bg.jpg
│  │
│  ├─train
│  │   a4d2cef8-hutao_bg.jpg
│  │
│  └─val
│      a4d2cef8-hutao_bg.jpg
│
└─labels
    ├─test
    │   a4d2cef8-hutao_bg.txt
    │
    ├─train
    │   a4d2cef8-hutao_bg.txt
    │
    └─val
        a4d2cef8-hutao_bg.txt

• (Required) Create a data.yaml file in the root directory of the dataset. The content and directory structure are as follows:

Note: In the data.yaml file, nc represents the total number of label categories in the dataset, and names refers to the names of these labels. This information can be found in the classes.txt file.

File contents

# dataset path
train: ./images/train
val: ./images/val
test: ./images/test

# number of classes
nc: 3

# class names
names: [
  "bag",
  "hutao",
  "other person",
]

Directory structure

.
│ classes.txt
│ data.yaml
│ notes.json
│
├─images
│  ├─test
│  │   a4d2cef8-hutao_bg.jpg
│  │
│  ├─train
│  │   a4d2cef8-hutao_bg.jpg
│  │
│  └─val
│      a4d2cef8-hutao_bg.jpg
│
└─labels
    ├─test
    │   a4d2cef8-hutao_bg.txt
    │
    ├─train
    │   a4d2cef8-hutao_bg.txt
    │
    └─val
        a4d2cef8-hutao_bg.txt

With this, a custom dataset suitable for YOLOv8 has been successfully created. In our next issue, we will explain how to train a custom dataset using YOLOv8(Training/inference on a custom dataset in Label Studio using YOLOv8)。.

• AliPay
• WeChat