ABIS Finger Things

Known External Benchs

We can go with FVC-STD-1.0 here.

https://biolab.csr.unibo.it/fvcongoing/UI/Form/BenchmarkAreas/BenchmarkAreaFV.aspx Although this is a 1-1 benchmark suite. This is a good reference for point for accuracy benchs.

NIST's evals

This contains multiple evaluations, but we can focus on 1-N matching here. https://github.com/usnistgov/frifte

Test sets

We need to test how good our system is at narrowing down the search space. This can be part of our evaluation sets that we have built.

Ideas for Indexing

Locality Sensitive Hashing (LSH): Reduces dimensionality while preserving similarity relationships

How does this work

Multi-level hierarchical matching: Initial coarse filtering followed by refined matching

what are the hierarchies

An efficient minutiae based geometric hashing for fingerprint database

local doc copy

Quality Check Ideas

Cool Papez

Learning a Fixed-Length Fingerprint Representation

this paper looks like a well researched paper and a landmark in the fingerprint indexing in the CNN scene.

Implementation Plans

LSH based indexing

I want to improve the indexing mechanism of my AFIS, these are the evaluation metrics that I've considered for building the system. This seem about right to me but I'm sure I might have missed some metric. This will be used to evaluate the different indexing algorithms and implementations. Let me know what I have missed.

Evaluation Metrics

  • Penetration Rate
  • Dataset size (i.e: number of fingerprints)
  • DB/Storage required to represent those record
  • Storage Growth rate
  • Memory Required to represent one of that record
  • Memory growth rate
  • Search Response Time(i.e: time to compare 1 vs N records)
  • Search Response Time Growth Rate
  • Computation Power Requirements
  • Accuracy Measurements
    • TPR (Hit Rate)
    • FPR (False Hit Rate)
    • Miss rate (False Negative Rate)
    • Recall(i.e: TPR / (TPR + FNR))
    • Do we need CMC curves? I've seen it being used in the evaluations not quite sure if its required for me.
  • Insertion Metrics
    • Enrolment Time / Throughput
    • Index-Build Time