| # Technical Documentation for the Text-to-Video Dataset “VidData” |
|
|
| ## 1. Introduction |
| This dataset contains 1006 annotated videos of everyday scenes, used for training and evaluating AI models in video generation and recognition. It is structured to meet the needs of Text-to-Video models and motion analysis. |
|
|
| ## 2. Dataset Specifications |
|
|
| ### 2.1. Generation Criteria |
| - **Maximum video duration**: 10 seconds maximum |
| - **Video themes**: |
| - Walking |
| - Exercising |
| - Writing |
| - Shopping |
| - Sleeping |
| - Meditating |
| - Working |
| - Studying |
| - Driving |
| - Washing |
| - Gardening |
| - Calling |
| - Listening |
| - Organizing |
| - Planning |
| - Relaxing |
| - Teaching |
| - **Video size**: 512×512 pixels |
|
|
| ### 2.2. Dataset Organization |
| The dataset is organized under a main folder called VidData, which includes three essential parts: |
| data/train/: Contains a VidData.csv file, likely storing metadata or structured details about the videos. |
| video/: Holds the video files (e.g., ---_iRTHryQ_13_0to241.mp4), named in a specific format, possibly indicating segments or unique identifiers. |
| readme.md: Provides documentation about the dataset's structure and usage. |
| This structure clearly separates raw video data, metadata (CSV), and documentation, ensuring efficient organization for analysis and processing. |
| data/train/: Contains CSV files with video-related metadata. |
| video/: Stores the actual video files. |
| ## 3. Data Structure |
| The dataset is stored as a CSV file and includes the following columns: |
| |
| | Column | Type | Description | |
| |-------------------------|---------|--------------------------------------| |
| | video | string | Video file name | |
| | caption | string | Textual description of the video | |
| | temporal consistency score | float64 | Temporal consistency score | |
| | fps | float64 | Frame per second | |
| | frame | int64 | Number of frames in the video | |
| | seconds | float64 | Video duration in seconds | |
| | motion score | float64 | Motion score | |
| | camera motion | string | Type of camera motion (e.g., pan_left) | |
| |
| ## 4. Libraries Used |
| |
| ### 4.1. Library Examples |
| Here are some example libraries that can be used when analyzing this data: |
| - **OpenCV**: Video manipulation and processing (reading, writing, frame extraction, contour detection, filtering, etc.). |
| - **Scikit-Image**: Calculating the Structural Similarity Index (SSIM) for image quality evaluation and various image transformations (segmentation, filtering, etc.). |
| - **NumPy**: Efficient manipulation of matrices and arrays, essential for calculations on images and videos. |
| - **Pandas**: Managing and structuring metadata associated with videos (e.g., file names, timestamps, annotations). |
| - **Matplotlib/Seaborn**: Visualizing analysis results as graphs. |
| |
| ### 4.2. Installing Dependencies |
| Follow the instructions below to install the required libraries: |
| 1. Create a `requirements.txt` file and add the following: |
| opencv-python==4.8.1.78 # Video manipulation and processing |
| scikit-image==0.22.0 # SSIM calculation and image transformations |
| numpy==1.26.2 # Efficient manipulation of matrices and arrays |
| pandas==2.1.4 # Managing and structuring metadata |
| matplotlib==3.8.2 # Visualizing analysis results |
| seaborn==0.12.2 # Advanced visualization with enhanced graphics |
| 2. Run the command: `pip install -r requirements.txt` |
| |
| **Note**: Only include the libraries you need in `requirements.txt`. |
| |
| ## 5. Using the Dataset |
| |
| ### 5.1. Primary Applications |
| #### 5.1.1. Text-to-Video Generation |
| - Train models to generate video based on textual input. |
| - Benchmark performance by comparing generated video against dataset entities. |
| |
| #### 5.1.2. Video Description Models |
| - Evaluate models designed to generate textual descriptions from videos. |
| |
| #### 5.1.3. Temporal Consistency Analysis |
| - Test model for maintaining smoothness and coherence in video generation. |
| |
| ### 5.2. Example Workflow |
| - Load the dataset using Python: |
| ## 4. Libraries Used |
| ```python |
| import pandas as pd |
| |
| dataset = pd.read_csv('VidData.csv') |
| print(dataset.head()) |
|
|
| ##Access video metadata: |
| video = dataset.iloc[0] # First entry |
| print(f"Video Name: {video['video_name']}") |
| print(f"Caption: {video['Caption']}") |
| print(f"Duration: {video['duration_seconds']} seconds") |
|
|
| ##Filter video based on motion: |
| high_motion_videos = dataset[dataset['motion_score'] > 1.0] |
| print(high_motion_videos) |
| ``` |
| |
| ## 6. File Format |
| The dataset is delivered in CSV format, with each column representing a video and its metadata. |
| |
| ## 7. Sample Entry: |
| |
| | video\_name | caption | temporal\_consistency\_score | fps | frames | duration\_seconds | motion\_score | camera\_motion | |
| | ----------- | ------------------------------------------------------ | ---------------------------- | --- | ------ | ----------------- | ------------- | -------------- | |
| | E\_1.mp4 | The video shows a soccer player kicking a soccer ball. | 0.948826 | 30 | 195 | 6.5 | 0.826522 | 1.105807 | |
|
|
|
|
| ## 8. Contact |
| For inquiries, please contact: |
|
|
| - **Email**: [info@databoost.us](mailto\:info@databoost.us) |
| - **Website**: [databoost.us](https://databoost.us) |
