System and method for facial recognition accuracy

A digital data processing method of identifying a person of interest in a video stream, comprising executing on a digital data processing system that includes one or more digital data processing devices that are in communications coupling with one another the steps of

• A. detecting, using a detector and without using a facial recognizer, a face of a first individual in a first video frame of a video stream,
• B. using a tracker to track, without using a facial recognizer, faces associated with the first individual detected in multiple frames of the video stream,
• C. scoring, without using a facial recognizer, each of the detected faces,
• D. selecting, based on scores of the detected faces, faces in the multiple frames apparently associated with the first individual, and applying those selected faces to a facial recognizer,
• E. using an embedder to map each face selected in the selecting step to a corresponding vector of numbers,
• F. clustering the corresponding vectors of numbers into one or more clusters,
• G. using the one or more clusters to identify selected faces associated with different individuals.

A digital data processing method of identifying a person of interest in a video stream comprising executing, on a digital data processing system that includes one or more central processing units (CPUs) that are coupled to one or more graphics processing units (GPUs), the steps of

• A. detecting, using a detector and without using a facial recognizer, a face of a first individual in a first video frame of a video stream,
• B. using a tracker to track, without using a facial recognizer, faces associated with the first individual detected in multiple frames of the video stream,
• C. scoring, without using a facial recognizer, each of the detected faces,
• D. selecting based on scores of the detected faces in the multiple frames apparently associated with the first individual, and applying those selected faces to a facial recognizer,
• E. using an embedder to map each face selected in the selecting step to a corresponding vector of numbers,
• F. clustering the corresponding vectors of numbers into one or more clusters,
• G. using the one or more clusters to identify selected faces associated with different individuals,
• wherein, steps (A), (B) and (D) are executed on a said CPU and step (E) is executed on a said GPU.