Data Healthcheck for CV

The Dataset Health Check provides a detailed analysis of the dataset’s statistical view and quality metrics for the project’s objectives.

Note: After extracting the video, we obtain images. The same health check process outlined in Step show below should be applied to all the resulting data

💡To initiate a Data Health Check, follow these steps:

• Go to the Datasets.
• Click on “ View Details “.

• After Clicking on View Details, Select “ View Dataset HealthCheck Statistics”.

• Here the following pieces of information and Explained detailing how each works and why it is essential:

  • Number of Images : Total count of images available in the dataset folder.
  • Number of Annotations : Total count of annotations present in the dataset folder
  
  

  • Image Dimension Insights : Insights on image dimensions.

  • Class Distribution Histogram : Distribution of annotations across different classes.

  • Class Balance : Distribution of annotations across different classes, showing balance or skewness.

  • Outlier Metrics Summary : Provides a summary of data points that deviate significantly from the rest of the dataset, highlighting potential anomalies or errors.

  • Area score distribution : Measures the distribution of object sizes within images, indicating how well objects of varying sizes are represented.

  • Aspect Ratio score distribution : Evaluates the distribution of aspect ratios (width-to-height ratio) in images, ensuring diverse shape representation.

  • Blue Values score distribution : Analyzes the distribution of blue channel intensity values across images, assessing color balance and consistency.

  • Blur score distribution : Indicates the level of blur in images, assessing the clarity and focus quality of the dataset.

  • Brightness score distribution : Measures the distribution of brightness levels in images, helping to identify underexposed or overexposed content.

    

  • Contrast score distribution : Evaluates the distribution of contrast levels in images, determining the visual distinction between light and dark area.

    

  • Green Values score distribution : Analyzes the distribution of green channel intensity values, contributing to the overall color balance evaluation.

    

  • Image Diversity score distribution : It provides an understanding of how distinct each image is relative to the others, ensuring that the dataset captures a wide range of scenarios, patterns, or characteristics.

    

  • Image Singularity score distribution : Measures how unique or distinctive each image is compared to the rest, identifying potential duplicates or overly similar content.

    

  • Random Values on Images score distribution : Evaluates the distribution of randomly selected pixel values, serving as a control metric for noise or unintended patterns.

    

  • Red Values score distribution : Analyzes the distribution of red channel intensity values in images, contributing to the assessment of color balance.

    

  • Sharpness score distribution : Measures the distribution of image sharpness, indicating the level of detail and clarity across the dataset

💡 Note: IF you observe class imbalance in your dataset, where one class has a significantly higher count than the other (for example, in a two-class dataset, one class is overrepresented), it may lead to biased model training and reduced performance. To mitigate this, we have an option called +Generate. This will help ensure your dataset is properly balanced, minimizing the risks associated with skewed class distribution.

💡 Steps for “Generate +“

• Before Clicking on “ Generate”, Go Back to Dataset.
• Upload Raw Images as per your requirements.
  • Use “Auto Annotate”to automatically annotate the newly uploaded images, or “Manually Add” the JSON and image folders if annotations already exist.
• Click “Generate”
  • This will trigger the dataset balancing process.
• “View” Updated Health Statistics
  • After the balancing process is completed, the updated health statistics will be displayed, reflecting the adjustments made for class distribution
• Save the newly balanced dataset. It is now ready for Visualization.