VeriLook 10.0 SDK Feedback

General information about Neurotechnology and its software

VeriLook 10.0 SDK Feedback

Postby win32api » Jun 21, 2017 19:19

Good day.
I downloaded the new version (10.0) and tested it on our database.
The result frightened me. Very very bad. The new version can not find a face on very clear photos. 30% of our photos were rejected, while version 9.0 without any problems identifies the face and marks all the points correctly. Almost half of the complex photographs (poorly visible face, head inclinations, etc.) the new version began to determine the face is not correctly. At the same time, version 9.0 easily copes with them.
Tried to change the settings (although on version 9.0 everything worked well and without any settings) of FacesConfidenceThreshold and FacesLivenessThreshold but nothing helped.
Will these problems be resolved in the near future??? We can not conduct a full-fledged test, because half of the photographs have been determined incorrectly. The new version strongly digs through the previous one.
win32api
 
Posts: 9
Joined: Aug 01, 2016 21:28

Re: VeriLook 10.0 SDK Feedback

Postby MartynasV » Jun 22, 2017 12:04

Hello win32api,

That is strange. Try lowering biometricClient.FacesQualityThreshold, at first set it to zero and test how it impacts results.
Additionally you could try setting "Faces.MaximalYaw" property to 90 (biometricClient.SetProperty("Faces.MaximalYaw",90)). Try these two and let me know.
Martynas V
Neurotechnology
MartynasV
Neurotechnology
 
Posts: 258
Joined: Feb 27, 2015 11:24
Location: Vilnius, Lithuania

Re: VeriLook 10.0 SDK Feedback

Postby win32api » Jun 22, 2017 17:48

MartynasV wrote:Hello win32api,

That is strange. Try lowering biometricClient.FacesQualityThreshold, at first set it to zero and test how it impacts results.
Additionally you could try setting "Faces.MaximalYaw" property to 90 (biometricClient.SetProperty("Faces.MaximalYaw",90)). Try these two and let me know.


Hello MartynasV,

Yes, we tried to change and calibrate the settings - it does not help.
I created a collage for comparison. For testing, I just took random photos from the internet.
I started the demo program that came with the library (SimpleFacesSampleCS) and on the "GeneralizeFaces" tab I added only 2 settings:
Code: Select all
BiometricClient.FacesMaximalRoll = 180;
BiometricClient.FacesMaximalYaw = 90;

Here is the result of comparing the two versions: http://i.imgur.com/i42dShb.jpg

Further on the "Identify Face" tab I set the following settings (we use the same settings in our system for processing complex photos and everything works well):
Code: Select all
BiometricClient.FacesMaximalRoll = 180;
BiometricClient.FacesMaximalYaw = 90;
BiometricClient.FacesConfidenceThreshold = 0;
BiometricClient.FacesQualityThreshold = 0;
BiometricClient.FacesSharpnessThreshold = 0;
BiometricClient.FacesLivenessThreshold = 20;

Here is the result of comparing the two versions: http://i.imgur.com/ORqRQAk.jpg

There are a lot of such examples, as I already said almost 20-30% of the photos in our database are not determined correctly, although everything was fine in version 9.0.
We hope that these problems will be corrected, because we can not switch to the new version.
win32api
 
Posts: 9
Joined: Aug 01, 2016 21:28

Re: VeriLook 10.0 SDK Feedback

Postby MartynasV » Jun 23, 2017 14:40

win32api wrote:
MartynasV wrote:Hello win32api,

That is strange. Try lowering biometricClient.FacesQualityThreshold, at first set it to zero and test how it impacts results.
Additionally you could try setting "Faces.MaximalYaw" property to 90 (biometricClient.SetProperty("Faces.MaximalYaw",90)). Try these two and let me know.


Hello MartynasV,

Yes, we tried to change and calibrate the settings - it does not help.
I created a collage for comparison. For testing, I just took random photos from the internet.
I started the demo program that came with the library (SimpleFacesSampleCS) and on the "GeneralizeFaces" tab I added only 2 settings:
Code: Select all
BiometricClient.FacesMaximalRoll = 180;
BiometricClient.FacesMaximalYaw = 90;

Here is the result of comparing the two versions: http://i.imgur.com/i42dShb.jpg

Further on the "Identify Face" tab I set the following settings (we use the same settings in our system for processing complex photos and everything works well):
Code: Select all
BiometricClient.FacesMaximalRoll = 180;
BiometricClient.FacesMaximalYaw = 90;
BiometricClient.FacesConfidenceThreshold = 0;
BiometricClient.FacesQualityThreshold = 0;
BiometricClient.FacesSharpnessThreshold = 0;
BiometricClient.FacesLivenessThreshold = 20;

Here is the result of comparing the two versions: http://i.imgur.com/ORqRQAk.jpg

There are a lot of such examples, as I already said almost 20-30% of the photos in our database are not determined correctly, although everything was fine in version 9.0.
We hope that these problems will be corrected, because we can not switch to the new version.

Hello win32api,

We are unable to reproduce it. Try with lower BiometricClient.FacesMaximalRoll value than 180.
Are you running 64bit or 32bit?
Martynas V
Neurotechnology
MartynasV
Neurotechnology
 
Posts: 258
Joined: Feb 27, 2015 11:24
Location: Vilnius, Lithuania

Re: VeriLook 10.0 SDK Feedback

Postby win32api » Jun 23, 2017 14:54

Good day.
It's very easy to reproduce, I just run a demo project. I'm using x64 version. The fact is that with the same settings, the new version has become much worse to find the face. I can send you photos that I used during testing. They work fine for version 9, but for some reason they are poorly recognized in the 10th.
win32api
 
Posts: 9
Joined: Aug 01, 2016 21:28

Re: VeriLook 10.0 SDK Feedback

Postby MartynasV » Jun 26, 2017 12:22

win32api wrote:Good day.
It's very easy to reproduce, I just run a demo project. I'm using x64 version. The fact is that with the same settings, the new version has become much worse to find the face. I can send you photos that I used during testing. They work fine for version 9, but for some reason they are poorly recognized in the 10th.

Hello win32api,

I found image with Johnny Depp that you have used and was able to reproduce the issue. If you try to detect face in image using "Detect faces" tab from SimpleFacesSample, then it would detect two faces(one correct, one false) - by default our algorithm favors bigger faces and that incorrectly detected face is bigger, so when you perform extraction(in "Identify" tab) only one face is detected(the bigger one). To fix that you can change "Faces.FavorLargestFace" property to false(_biometricClient.SetProperty("Faces.FavorLargestFace", false);), that way face with bigger confidence will be used.
Martynas V
Neurotechnology
MartynasV
Neurotechnology
 
Posts: 258
Joined: Feb 27, 2015 11:24
Location: Vilnius, Lithuania

Re: VeriLook 10.0 SDK Feedback

Postby win32api » Jun 26, 2017 12:39

Yes! Excellent, with this setting everything fell into place. Now everything works well, thank you very much! :)
win32api
 
Posts: 9
Joined: Aug 01, 2016 21:28

Re: VeriLook 10.0 SDK Feedback

Postby win32api » Jun 26, 2017 22:51

It might be better for you to make this option disabled by default? Because with it library begins to make mistakes even on very simple photos and finds a face in unpredictable places
win32api
 
Posts: 9
Joined: Aug 01, 2016 21:28

Re: VeriLook 10.0 SDK Feedback

Postby win32api » Jun 27, 2017 20:38

With the disabling of this setting everything has become much much better, BUT finding a face is still worse than the previous version. We have decided not to upgrade to the new version for a while. I hope this will improve with time, because we ourselves are interested in this. Thanks for the good product, good luck to you.
win32api
 
Posts: 9
Joined: Aug 01, 2016 21:28

Re: VeriLook 10.0 SDK Feedback

Postby MartynasV » Jun 28, 2017 08:46

win32api wrote:With the disabling of this setting everything has become much much better, BUT finding a face is still worse than the previous version. We have decided not to upgrade to the new version for a while. I hope this will improve with time, because we ourselves are interested in this. Thanks for the good product, good luck to you.

Hello win32api,

Our face detection wasn't changed in 10.0 version, but when you set "FacesMaximalYaw" value higher than 45, then our algorithm uses new face detection, which is able to detect face in profile position. If you don't need than then you should set Maximal Yaw value lower than 45 and old detection(same as in 9.0) will be used.
Martynas V
Neurotechnology
MartynasV
Neurotechnology
 
Posts: 258
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