The Head Pose demonstration showcases a detecting faces and reporting where they are facing. It uses a combination of a model trained to detect faces along with another model trained to estimate the head pose (yaw, pitch, roll). The results are graphically overlaid over the live camera feed.
The application includes a metrics overlay showing the camera, display, and inference visual statistics and timing details.
The Head Pose application contains two different models running: the face detection model and a head-pose model. Statistics for each model will be noted separately where applicable. Because there are two models running, overall inference FPS is impacted.
Reported FPS and Resolutions
|Item||Frames per Second||Resolution|
512x384 (face detection)
224x224 (head pose)
Inference Timing Details
|Item||Timing, Face Detection (ms)||Timing, Head Pose (ms)|
|Time to load the frame into the model||2.31||4.78|
|Running model inference on the frame||11.53||14.84|
|Decode head pose||2.22||0.05|