3 Metrics

3.1 Location

When a metric depicts search accuracy in this document, it is reported in terms of Location: Region and Subject.

  • Region: The correct region of the correct subject was returned.
    • For search probes sourced from a distal phalanx (i.e., a “latent fingerprint”), the correct finger position 1–10 shall be returned.
    • For search probes sourced from a palm or a non-distal phalanx, the most localized region shall be returned. Some palm regions may be interchangeable based on the exemplars provided (e.g., a palm probe’s source could reasonably be seen in a lower palm, hypothenar, and writer’s palm exemplar). Credit is given for Region in this case.
  • Subject: Any position from the correct subject is returned. This is designed to reward the situation where an implementation cannot ascertain the most localized region from the set of exemplars enrolled and may indicate segmentation error.

3.2 Cumulative Match Characteristic (CMC)

The Cumulative Match Characteristic (CMC) plots in this document show the false negative identification rate (FNIR) without respect for similarity score when searching probes against a enrollment database where a single mated identity for each search probe was present.

  •  ≈ 1 600 000 non-mated subjects were enrolled.
    • All subjects had at least one, but typically twenty, images containing distal phalanges. This typically included ten individually rolled impressions and “Identification Flat” captures featuring more than one distal phalanx per image that must be segmented by the implementation.
    •  ≈ 150 000 had one or more palm captures to enroll.
  • The requested size of the candidate list was always 100 subjects.
  • All possible Extended Feature Set (EFS) data was provided when “Image + EFS” is listed for probes. The type and quantity of EFS data present varies for each sample in each dataset and may have been entirely omitted. Initial experiments show nominal (if any) change when EFS data was provided alongside exemplars.
  • Probe impression type was always “Unknown Finger” or “Unknown Palm,” as appropriate. Future studies may show results using the impression type “Unknown Friction Ridge” for both types of probes.
  • The metric hit rate is equivalent to 1 − miss rate, or 1 − FNIR. For example, an FNIR of 0.1 indicates a hit rate of 0.9 (i.e., 90%).

3.3 Detection Error Tradeoff (DET)

The Detection Error Tradeoff (DET) plots in this document show the tradeoff between false positive and false negative identification rates when searching probes against a enrollment database where a single mated identity for each search probe was present.

  •  ≈ 1 600 000 non-mated subjects were enrolled.
    • All subjects had at least one, but typically twenty, images containing distal phalanges. This typically included ten individually rolled impressions and “Identification Flat” captures featuring more than one distal phalanx per image that must be segmented by the implementation.
    •  ≈ 150 000 had one or more palm captures to enroll.
    • Non-mated similarity scores come from rank = 1 when searching probes against an enrollment dataset without any mated subjects enrolled.
  • The requested size of the candidate list was always 100 subjects.
    • Mated similarity scores come from the correct location appearing at any rank.
  • All possible EFS data was provided when “Image + EFS” is listed for probes. The type and quantity of EFS data present varies for each sample in each dataset and may have been entirely omitted. Initial experiments show nominal (if any) change when EFS data was provided alongside exemplars.
  • Probe impression type was always “Unknown Finger” or “Unknown Palm,” as appropriate. Future studies may show results using the impression type “Unknown Friction Ridge” for both types of probes.