Muokkaa

Jaa


Face recognition data structures

This article explains the data structures used in the Face service for face recognition operations. These data structures hold data on faces and persons.

Caution

Face service access is limited based on eligibility and usage criteria in order to support our Responsible AI principles. Face service is only available to Microsoft managed customers and partners. Use the Face Recognition intake form to apply for access. For more information, see the Face limited access page.

Data structures used with Identify

The Face Identify API uses container data structures to the hold face recognition data in the form of Person objects. There are three types of containers for this, listed from oldest to newest. We recommend you always use the newest one.

PersonGroup

PersonGroup is the smallest container data structure.

  • You need to specify a recognition model when you create a PersonGroup. When any faces are added to that PersonGroup, it uses that model to process them. This model must match the model version with Face ID from detect API.
  • You must call the Train API to make any new face data reflect in the Identify API results. This includes adding/removing faces and adding/removing persons.
  • For the free tier subscription, it can hold up to 1000 Persons. For S0 paid subscription, it can have up to 10,000 Persons.

PersonGroupPerson represents a person to be identified. It can hold up to 248 faces.

Large Person Group

LargePersonGroup is a later data structure introduced to support up to 1 million entities (for S0 tier subscription). It is optimized to support large-scale data. It shares most of PersonGroup features: A recognition model needs to be specified at creation time, and the Train API must be called before use.

Person Directory

PersonDirectory is the newest data structure of this kind. It supports a larger scale and higher accuracy. Each Azure Face resource has a single default PersonDirectory data structure. It's a flat list of PersonDirectoryPerson objects - it can hold up to 20 million.

PersonDirectoryPerson represents a person to be identified. Updated from the PersonGroupPerson model, it allows you to add faces from different recognition models to the same person. However, the Identify operation can only match faces obtained with the same recognition model.

DynamicPersonGroup is a lightweight data structure that allows you to dynamically reference a PersonDirectoryPerson. It doesn't require the Train operation: once the data is updated, it's ready to be used with the Identify API.

You can also use an in-place person ID list for the Identify operation. This lets you specify a more narrow group to identify from. You can do this manually to improve identification performance in large groups.

The above data structures can be used together. For example:

  • In an access control system, The PersonDirectory might represent all employees of a company, but a smaller DynamicPersonGroup could represent just the employees that have access to a single floor of the building.
  • In a flight onboarding system, the PersonDirectory could represent all customers of the airline company, but the DynamicPersonGroup represents just the passengers on a particular flight. An in-place person ID list could represent the passengers who made a last-minute change.

For more details, please refer to the PersonDirectory how-to guide. A quick comparison between LargePersonGroup and PersonDirectory:

Detail LargePersonGroup PersonDirectory
Capacity A LargePersonGroup can hold up to 1 million PersonGroupPerson objects. The collection can store up to 20 millions PersonDirectoryPerson identities.
PersonURI /largepersongroups/{groupId}/persons/{personId} (/v1.0-preview-or-above)/persons/{personId}
Ownership The PersonGroupPerson objects are exclusively owned by the LargePersonGroup they belong to. If you want a same identity kept in multiple groups, you will have to Create Large Person Group Person and Add Large Person Group Person Face for each group individually, ending up with a set of person IDs in several groups. The PersonDirectoryPerson objects are directly stored inside the PersonDirectory, as a flat list. You can use an in-place person ID list to Identify From Person Directory, or optionally Create Dynamic Person Group and hybridly include a person into the group. A created PersonDirectoryPerson object can be referenced by multiple DynamicPersonGroup without duplication.
Model The recognition model is determined by the LargePersonGroup. New faces for all PersonGroupPerson objects will become associated with this model when they're added to it. The PersonDirectoryPerson object prepares separated storage per recognition model. You can specify the model when you add new faces, but the Identify API can only match faces obtained with the same recognition model, that is associated with the query faces.
Training You must call the Train API to make any new face/person data reflect in the Identify API results. There's no need to make Train calls, but API such as Add Person Face becomes a long running operation, which means you should use the response header "Operation-Location" to check if the update completes.
Cleanup Delete Large Person Group will also delete the all the PersonGroupPerson objects it holds, as well as their face data. Delete Dynamic Person Group will only unreference the PersonDirectoryPerson. To delete actual person and the face data, see Delete Person.

Data structures used with Find Similar

Unlike the Identify API, the Find Similar API is designed to be used in applications where the enrollment of Person is hard to set up (for example, face images captured from video analysis, or from a photo album analysis).

FaceList

FaceList represent a flat list of persisted faces. It can hold up 1,000 faces.

LargeFaceList

LargeFaceList is a later version which can hold up to 1,000,000 faces.

Next steps

Now that you're familiar with the face data structures, write a script that uses them in the Identify operation.