Reidentification
Identity across views, voices and sessions.
A live camera-and-audio system combines person, body, face and voice information. Learned profiles reconnect observations to a stored identity rather than relying only on temporary tracker or speaker labels.
01 / The interaction
How can recognition survive a new session?
The system keeps learned face and voice profiles, names and their associations across sessions. A new speaker label can change after a restart; matching stored voice features reconnects it with the existing identity.
Temporary tracking labels and persistent identities are separate concepts.
02 / Research & engineering
What went into it.
Visual recognition
Person detection and body/face recognition supply visual identity cues from the camera stream.
Voice profiles
Diarization and voice-fingerprint extraction support speaker recognition and stored profile matching.
Identity fusion
Fusion logic combines cues and maintains associations between voice profiles and people.
GPU execution
Six neural-model components run on CUDA. Capture, mouth-motion processing, fusion and the UI remain on CPU.
Compatibility
CPU/CUDA feature comparisons and saved-name recovery checks helped preserve existing profiles during the migration.
03 / Progress
Built. Tested. Next.
Six GPU model components
The live application reported all six neural-model components using CUDA.
Stored-profile compatibility
Existing names and associations were preserved and recovered after the inference migration.
81 → 56 ms per frame
A recorded local camera benchmark used about 31% less processing time. This describes that test setup, not general throughput.
75 software tests
The migration included software checks plus CPU/CUDA feature-compatibility tests.
The recorded benchmark evaluates processing time on the tested setup. It does not establish recognition accuracy across populations or operating conditions.
Continue exploring