Plumerai Familiar Face Identification¶

At Plumerai we’ve built the world’s most accurate familiar face identification that is tiny and fast. It identifies people with a high accuracy in any situation, and runs at the edge or in the cloud.

See live demos and a high-level description on our main Familiar Face Identification web page.

For technical documentation, please see the following sections:

For information about tracking, robustness, camera support, privacy, and lighting, distance, mounting, and location requirements, see the object detection page.

We provide two versions of familiar face identification:

Version 1: With automatic face enrollment¶

This version automatically enrolls new faces in the face library. The user can tag each identified person with a name afterward.

Example of Familiar Face Identification with automatic face enrollment where a person is first identified as a stranger, followed by the user labeling herself as 'Yara'.

Try out our automatic enrollment face identification live in the browser with your webcam.

Version 2: With manual face enrollment¶

This version allows the user to explicitly enroll new faces in the face library. This is typically used in a scenario where users scan all sides of their face with an app on a phone.

Example of Familiar Face Identification with manual face enrollment where a person named 'Yara' is explicitly enrolling herself, and then later in another scene is re-identified as herself.

You can try our manual enrollment face identification live in the browser with your webcam.

Familiar face identification features¶

State-of-the-art face identification¶

Our solution uses an end-to-end deep learning approach that consists of three neural networks: one for object detection, one for face representation, and one for face matching. We have applied various advanced model design, compression, and training techniques to make these networks fit within the hardware constraints of small CPUs and NPUs, while retaining excellent accuracy.

Familiar Face Identification examples — Familiar Face Identification in practice.

Enrollment¶

Our algorithm supports loading and storing a face library. The storage requirement for an enrollment of a single user is very small, as low as 520 bytes. For ease of integration, the face enrollment API is integrated in the regular usage API. In other words, Plumerai provides a single library/API that can both be used for regular operation and adding new faces to the face library, which makes it easy for developers to start using the Plumerai software.

Identity assignment¶

Our algorithms assign identities to person boxes, but it does so using face detections. When it can uniquely match a face to a person, the identity of the person box should match that of the matched face box. However, the algorithm can’t always be certain which face belongs to which tracked person, especially when people are standing close to each other. Tracking people is also more reliable than tracking faces, since people’s faces can for instance be turned away from the camera for a moment. Therefore, it is common that the identity of the person is known even when the identity of a face is unknown. Also, a new face might be associated with an already identified person. The identity of that person will then be determined by the information of both faces that were associated with it, and the strongest match will be selected.

Stranger detection¶

Familiar face identification can also be used to detect strangers. The software can distinguish between people in the face library, people that are not clearly visible, and people that are definitely not in the face library. The last category can be used to trigger an alarm or other action.