Image Quality Metrics Guidance

Choosing to import images that have been flagged as “low-quality” will cause more false positives to occur as SAFR incorrectly identifies newly scanned faces as identical to the low-quality facial image. Greater discrepancies between the recommended metric value and the actual metric value will result in more false positives. Similarly, having more than one metric value be poor or very poor will also result in more false positives.

Center Pose

Center Pose = .89 Center Pose = .76 Center Pose = .54 Center Pose = .34 Center Pose = .21

Center pose represents how directly the face is looking at the camera. The more the face looks up, down, left, or right of the camera, the more this metric value is reduced from 1. Similarly, if the face is tilted in any way (e.g. the person’s chin is pointing at a corner of the image) this metric value is reduced. The default recommended minimum value for this metric is .59. You can adjust the recommended minimum value by going to Tools -> Preferences, clicking on the Recognition tab, then adjusting the For merging slider in the Minimum required center pose quality section.

Quality Label Metric Range Description
Excellent 0.7 - 1.0 Full recognition accuracy can be expected under all conditions.
Good 0.6 - 0.7 Very good recognition accuracy can be expected in general but may confuse closely related family members.
Marginal 0.45 - 0.6 Good recognition but may result in occasional failures.
Poor 0.3 - 0.45 Recognitions can be performed to significant extent but may produce false recognitions.
Very Poor 0.0 - 0.3 Recognitions can still be performed but with significant possibility of confusing similar faces.

Sharpness

Sharpness = .79 Sharpness = .62 Sharpness = .58 Sharpness = .35 Sharpness = .22

Sharpness represents how clear the facial image is. The more blurry the face is, the more this metric value is reduced from 1. The default recommended minimum value for this metric is .45. You can adjust the recommended minimum value by going to Tools -> Preferences, clicking on the Recognition tab, then adjusting the For merging slider in the Minimum required face sharpness quality section.

Quality Label Metric Range Description
Excellent 0.7 - 1.0 Full recognition accuracy can be expected under all conditions.
Good 0.6 - 0.7 Very good recognition accuracy can be expected in general but may confuse closely related family members.
Marginal 0.45 - 0.6 Good recognition but may result in occasional failures.
Poor 0.3 - 0.45 Recognitions can be performed to significant extent but may produce false recognitions.
Very Poor 0.0 - 0.3 Recognitions can still be performed but with significant possibility of confusing similar faces.

Contrast

Contrast = 1 Contrast = .87 Contrast = .63 Contrast = .47 Contrast = .40 Contrast = .20

Contrast represents the color contrast within the facial image. The less color contrast a face has, the more this metric value approaches 0. The default recommended minimum value for this metric is .45. You can adjust the recommended minimum value by going to Tools -> Preferences, clicking on the Recognition tab, then adjusting the For merging slider in the Minimum required face contrast quality section.

Quality Label Metric Range Description
Excellent 0.7 - 1.0 Full recognition accuracy can be expected under all conditions.
Good 0.6 - 0.7 Very good recognition accuracy can be expected in general but may confuse closely related family members.
Marginal 0.45 - 0.6 Good recognition but may result in occasional failures.
Poor 0.3 - 0.45 Recognitions can be performed to significant extent but may produce false recognitions.
Very Poor 0.0 - 0.3 Recognitions can still be performed but with significant possibility of confusing similar faces.

Face Size

Face size defines the minimum required face size in pixels. The metric also includes a margin around the face. The margin is required when learning a face. The face itself (without the margin) includes the area ranging from the top of the forehead to the bottom of the chin and across the full width of the face excluding ears.

The recommended minimum value for this metric is 220 pixels. You can adjust the recommended minimum value by going to Tools -> Preferences, clicking on the Recognition tab, then adjusting the For learning/strangers slider in the Minimum Required Face Size section.

Note that only the shortest side of the image is used for the purpose of determining the metric value. For example, a facial image that is 200 x 300 (including the margin) would be classified as Marginal, since the shortest side (200) falls in the Marginal range.

Quality Label Metric Value Description
Excellent 260 px and greater Full recognition accuracy can be expected under all conditions.
Good 210 px - 260 px Very good recognition accuracy can be expected in general but may confuse closely related family members.
Marginal 160 px - 210 px Good recognition but may result in occasional failures.
Poor 110 px - 160 px Recognitions can be performed to significant extent but may produce false recognitions of blurry or otherwise not clearly visible faces.
Very Poor 60 px - 110 px Recognitions can still be performed but with significant possibility of confusing similar faces.

Occlusion

Occlusion represents how much of the face is occluded. Faces can be occluded by masks, baseball caps, or even the person’s hands held between the face and the camera. The default recommended maximum value for this metric is .5. You can adjust the recommended maximum value by going to Tools -> Preferences, clicking on the Recognition tab, then adjusting the For learning/strangers slider in the Maximum allowed occlusion section.

Quality Label Metric Range Description
Occluded 0.5 - 1.0 At least one of the facial features is not clearly visible thus potentially preventing full recognition accuracy. Recognition based on occluded features will not be possible and incorrect recognition of similar faces occluded in similar manner is possible.
Recognition is generally possible as long as two out of three key features (eyes, nose, mouth) are visible.
Not Occluded 0.0 - 0.5 All facial features are clearly visible and full recognition accuracy can be achieved.

Sentiment

Sentiment represents how happy (a positive sentiment score) or angry (a negative sentiment score) a face is. 0 sentiment (a neutral or serious expression) yields the most accurate facial recognition.