Optimize Face Analysis Performance

The quality of the image or video that you send to Media Server can have a significant effect on the performance of face analysis.

  • To be detected, faces must exceed the minimum size specified by the MinSize parameter in your task configuration.
  • To recognize faces and run demographics analysis and state analysis, the face must be large enough in the image that facial features are clearly visible.
  • Faces should be in focus, not blurry.
  • Face detection performs best when faces are looking towards the camera, so that both eyes are visible, but faces can be detected when viewed in profile (side-on). For face recognition, demographics analysis, and expression analysis, the person must be looking toward the camera so that both eyes are visible.
  • Ideally faces should be fully visible and not be partially obscured (for example a person standing behind a microphone).
  • Although face detection can process a relatively wide range of facial expressions, faces with neutral expressions are usually detected with the greatest reliability. Particularly unusual facial expressions might cause face detection and recognition to fail.
  • Spectacles or large amounts of facial hair increase the difficulty in detecting and recognizing faces and accuracy may be reduced in these cases.
  • The image or video should have even illumination, because dark shadows or bright highlights on the face reduce accuracy.
  • The image or video should not be affected by significant compression artifacts that can affect some formats (such as highly compressed JPEG images). Micro Focus recommends that you do not save your image files as JPEG images with high levels of compression or transcode video that has been captured from a camera. If you are using a digital video recorder (DVR) to record the footage from a camera and are then sending the video to Media Server, ensure the DVR is saving the video at full resolution and is not reducing the bit rate.
  • For face recognition, Micro Focus recommends that you configure Media Server to return the top five or ten results and then have a person select the best match from these results. Using face recognition in this way can produce better accuracy than using Media Server alone.