Similarity Scores in Ranges

General information about Neurotechnology and its software

Similarity Scores in Ranges

Postby Ajmal Hamidi » Mar 27, 2017 07:19

Hello,

I was going through the MMA SDK recently and I faced the following question.

Reference - MMA SDK

Matching: The templates can be compared with the aim to check if they belong to the same person. The result of such comparison is the similarity score. The higher score suggests the higher probability that features collections are obtained from the same person. Most of the time the confidence score itself is not needed, rather the simple answer (are the templates from the same person or not). The score is mapped to yes/no answer with the matching threshold.

Question- Can Neurotech MMA provides similarity scores (or similarity %) in the following three ranges for duplication check;

A: from 0-35 range

B: from 36- 69 range

C: from 70-100 range

The reason for asking this question is that, it will be difficult for client to have engineers who can read/understand minutiae and take the final decision if a person is really dup/un-dup considering the FAR and FRR of MMA.

Regards,
Hamidi.
Ajmal Hamidi
 
Posts: 6
Joined: Jun 25, 2013 05:29

Re: Similarity Scores in Ranges

Postby MartynasV » Mar 27, 2017 12:02

Hello,

Question- Can Neurotech MMA provides similarity scores (or similarity %) in the following three ranges for duplication check;
A: from 0-35 range
B: from 36- 69 range
C: from 70-100 range

No, score range is 0-infinity. You can set matching threshold to 0 to get all results, then you can filter results by score as you wish.

About score and matching threshold.

The returned score should be interpreted as the probability that the false acceptance happened. The score is retuned by using such algorithm – if the matching score is equal or higher than the set matching threshold, then it means that modality has matched and score is returned. If the score is lower than the matching threshold, then “0” value is returned and it means that the modality did not match.
There is no maximum value for the matching score. Bigger the score lower the probability that false acceptance has happened. There is no maximum limit, because the bigger resolution image you match to itself the bigger the score will be returned.

Usually the matching threshold is set by the requirements of the project or by field tests. So we can not say that one or another threshold is good or not. You can find some additional information on matching threshold and score in the documentation of the SDK, Section “Matching threshold and score”.

Score as percentage.

This is quite a common request from our clients. However such representation in percentages as 90% or 100% would not have any mathematical meaning in relation of match or no match.

Returned score represent probability of false match of the particular template against probe template and can be calculated by formula: 10^(-Score/12). So in all cases (with a bit higher threshold) probability will be very small.
In case of 1 to N in addition it is important to take into account amount of records in database (DB) in which search was done. Probability of at least one false match in entire search will be equal to 1 – (1 – 10^(-score/12))^N. For more information please check documentation in Neurotec Biometrics SDK chapter 5.1 “Matching”.

For clarity to final customer you can select one score over which all matches will be counted as 100% match (lets say 150 or a bit more) and your current minimum threshold (for example 48) would be 0%. Then you could display percentage by this interval when probing your candidate templates.
Martynas V
Neurotechnology
MartynasV
Neurotechnology
 
Posts: 264
Joined: Feb 27, 2015 11:24
Location: Vilnius, Lithuania

Re: Similarity Scores in Ranges

Postby Ajmal Hamidi » Mar 29, 2017 04:41

Hello,

MartynasV

Thank you very much for the great details & explanation.
Ajmal Hamidi
 
Posts: 6
Joined: Jun 25, 2013 05:29

Re: Similarity Scores in Ranges

Postby nothingmore » Nov 27, 2017 09:56

Ajmal Hamidi wrote:Hello,

I was going through the MMA SDK recently and I faced the following question.

Reference - MMA SDK

Matching: The templates can be compared with the aim to check if they belong to the same person. The result of such comparison is the similarity score. The higher score suggests the higher probability that features collections are obtained from the same person. Most of the time the confidence score itself is not needed, rather the simple answer (are the templates from the same person or not). The score is mapped to yes/no answer with the matching threshold.

Question- Can Neurotech MMA provides similarity scores (or similarity %) in the following three ranges for duplication check;

A: from 0-35 range

B: from 36- 69 range

C: from 70-100 range

The reason for asking this question is that, it will be difficult for client to have engineers who can read/understand minutiae and take the final decision if a person is really dup/un-dup considering the FAR and FRR of MMA.

Regards,
Hamidi.


I want to get the all comparison scores in one text file? how can i do that with this software?
nothingmore
 
Posts: 3
Joined: Nov 27, 2017 09:35

Re: Similarity Scores in Ranges

Postby nothingmore » Nov 27, 2017 10:26

MartynasV wrote:Hello,

Question- Can Neurotech MMA provides similarity scores (or similarity %) in the following three ranges for duplication check;
A: from 0-35 range
B: from 36- 69 range
C: from 70-100 range

[url]No, score range is 0-infinity. You can set matching threshold to 0 to get all results, then you can filter results by score as you wish.[/url]

About score and matching threshold.

The returned score should be interpreted as the probability that the false acceptance happened. The score is retuned by using such algorithm – if the matching score is equal or higher than the set matching threshold, then it means that modality has matched and score is returned. If the score is lower than the matching threshold, then “0” value is returned and it means that the modality did not match.
There is no maximum value for the matching score. Bigger the score lower the probability that false acceptance has happened. There is no maximum limit, because the bigger resolution image you match to itself the bigger the score will be returned.

Usually the matching threshold is set by the requirements of the project or by field tests. So we can not say that one or another threshold is good or not. You can find some additional information on matching threshold and score in the documentation of the SDK, Section “Matching threshold and score”.

Score as percentage.

This is quite a common request from our clients. However such representation in percentages as 90% or 100% would not have any mathematical meaning in relation of match or no match.

Returned score represent probability of false match of the particular template against probe template and can be calculated by formula: 10^(-Score/12). So in all cases (with a bit higher threshold) probability will be very small.
In case of 1 to N in addition it is important to take into account amount of records in database (DB) in which search was done. Probability of at least one false match in entire search will be equal to 1 – (1 – 10^(-score/12))^N. For more information please check documentation in Neurotec Biometrics SDK chapter 5.1 “Matching”.

For clarity to final customer you can select one score over which all matches will be counted as 100% match (lets say 150 or a bit more) and your current minimum threshold (for example 48) would be 0%. Then you could display percentage by this interval when probing your candidate templates.


How can we save all results for the database?
is it possible to get an excel file or text file containing following information in each line:
*first image name* , *second image name*, and *similarity score*
?
nothingmore
 
Posts: 3
Joined: Nov 27, 2017 09:35

Re: Similarity Scores in Ranges

Postby MartynasV » Nov 27, 2017 11:44

Hello nothingmore,

How can we save all results for the database?
I want to get the all comparison scores in one text file? how can i do that with this software?
is it possible to get an excel file or text file containing following information in each line:
*first image name* , *second image name*, and *similarity score*?

There are no integrated tools in our SDK to save matching results nor to database, nor to excel file, nor to text file. You have to use standard libraries or 3rd party libraries to save those results to wanted medium.
Martynas V
Neurotechnology
MartynasV
Neurotechnology
 
Posts: 264
Joined: Feb 27, 2015 11:24
Location: Vilnius, Lithuania


Return to Public

Who is online

Users browsing this forum: No registered users and 3 guests