Skip to main content

Better Facial Recognition Clusters

Purpose​

This guide explains how to optimize facial recognition in systems with large image libraries. By following these steps, you'll achieve better clustering of faces, reducing the need for manual merging.

Important Notes​

  • Best Suited For: Large image libraries after importing a significant number of images.
  • Warning: This method deletes all previously assigned names.
  • Tip: Always take a backup before proceeding!

Step-by-Step Instructions​

Objective​

To enhance face clustering and ensure the model effectively identifies faces using qualitative initial data.

Steps​

1. Adjust Machine Learning Settings​

Navigate to:
Admin → Administration → Settings → Machine Learning Settings

Make the following changes:

  • Maximum recognition distance (Optional):
    Lower this value, e.g., to 0.4, if the library contains people with similar facial features.
  • Minimum recognized faces:
    Set this to a high value (e.g., 20 For libraries with a large amount of assets (~100K+), and 10 for libraries with medium amount of assets (~40K+)).

    A high value ensures clusters only include faces that appear at least 20/value times in the library, improving the initial clustering process.

2. Run Reset Jobs​

Go to:
Admin → Administration → Settings → Jobs

Perform the following:

  1. FACIAL RECOGNITION → Reset

These reset jobs rebuild the recognition model based on the new settings.

3. Refine Recognition with Lower Thresholds​

Once the reset jobs are complete, refine the recognition as follows:

  • Step 1:
    Return to Minimum recognized faces in Machine Learning Settings and lower the value to 10 (In medium libraries we will lower the value from 10 to 5).

    Run the job: FACIAL RECOGNITION → MISSING Mode

  • Step 2:
    Lower the value again to 3.

    Run the job: FACIAL RECOGNITION → MISSING Mode

try different values

For certain libraries with a larger or smaller amount of assets, other settings will be better or worse. It is recommended to try different values ​​before assigning names and see which settings work best for your library.