PaddleX Video Detection Task Module Data Annotation Tutorial

This document will guide you through the process of annotating data for video detection tasks.

1. Annotation

1.1 Data Preparation

Collect the raw video data and organize each video into different folders based on their categories. Supported video formats include '.mp4', '.avi', '.mov', '.mkv'. For example:

dataset_dir  # Root directory of the dataset; the directory name can be changed
├── class1    # Directory for storing videos of each category; the directory name can be changed, containing multiple video files
├── class2
├── ...

1.2 Video Data Conversion

The raw video data needs to be converted into image frames. You can use the convert_video_to_images.py script and run the following command:

python convert_video_to_images.py dataset_dir dataset_dir/rgb-images

Here, dataset_dir is the directory with videos to be annotated from section 1.1. rgb-images is the directory where the converted image frames will be saved, with each video's images stored in a separate subdirectory.

1.3 Image Frame Data Annotation

Annotation Process

• For annotation, you can refer to the Object Detection Annotation Documentation. Use Labelme or PaddleLabel to annotate the bounding boxes for each image, and save them in the coco.json format.

• After completing the annotation, you need to organize the annotation files into the format required in section 2. Data Format. Use the convert_coco_to_txt.py script to convert the coco.json annotation files into txt format. Place these files in the labels directory, with each annotation file corresponding to the image frames of a single video.

python convert_coco_to_txt.py coco.json dataset_dir

dataset_dir is the directory where you want to save the txt files. coco.json is the annotation file in coco format.

2. Data Format

• The directory structure for the final dataset should be as follows:

 dataset_dir    # Root directory of the dataset; the directory name can be changed
├── rgb-images     # Directory for saving video images; the directory name cannot be changed
│   ├── class1    # Directory for storing videos of each category; the directory name can be changed
│   │   ├── video1_images  # Directory for storing images from each video in the category; the directory name can be changed
│   │   │   ├── 00001.jpg  # Image naming rule: frame number.jpg, e.g., 00001.jpg
│   │   │   ├── 00002.jpg
│   │   ├── video2_images
│   ├── class2
│   ├── ...
├── labels       # Directory for saving labels of the video images; the directory name cannot be changed and corresponds to the subdirectory names in rgb-images
│   ├── class1    # Directory for storing video labels of each category; the directory name can be changed
│   │   ├── video1_images  # Directory for storing image labels from each video in the category; the directory name can be changed
│   │   │   ├── 00001.txt  # Each txt file corresponds to the label of an image, with each line formatted as: classid x1 y1 x2 y2
│   │   │   ├── 00002.txt
│   │   ├── video2_images
│   ├── class2
│   ├── ...
├── label_map.txt # File mapping label ids to category names; the file name cannot be changed. Each line provides a category id and name, e.g., 1 Basketball
├── train.txt     # Training set annotation file, with each line providing the path to a video image label, e.g., labels/Biking/v_Biking_g02_c01/00228.txt
└── val.txt       # Validation set annotation file, with each line providing the path to a video image label

The annotation files use an image format. Please refer to the above specifications to prepare your data. Additionally, you can refer to the example dataset for guidance.