photools.com Community

Hi all,

I was playing around a little bit with IMatch face detection (not recognition) and I was comparing it with what other software is doing. My observations:

(1) IMatch is doing a very good job when comparing with most other software (DigiKam, Imageranger,...)
(2) I've observed that XnView MP - which has the ability of face detection via an add-on - is doing a little bit better.

If you want to test XnView MP (https://www.xnview.com/de/xnviewmp/) you can download it here (https://www.xnview.com/de/xnviewmp/) and the add-on (https://www.xnview.com/wiki/index.php?title=Face_recognition) here (https://www.xnview.com/wiki/index.php?title=Face_recognition). BUT: Automization is much much better in IMatch AND XnView MP has some serious bugs, especially XMP-metadata-writing is not working correctly. Therefore: Only use it with test-images!

I've seen in the XnView MP add-on folder that it seems to use OpenCV (opencv_world3416.dll). It seems that IMatch is also using OpenCV (looking at the Programs-folder).

Questions: Is there a setting in IMatch (haven't seen so fare) to improve face detection results? XnView MP has a "Threshold" setting (not sure what it is exactly doing).

In the two attachements you can see the different results from IMatch and from XnView MP. I've used the pictures from the txt-file (links in the file).

Thanks.

Quoteface detection via an add-on - is doing a little bit better.

How large is the sample size and variation you have used for your tests to make this statement?
Which settings have you used for the face recognition in IMatch?
By the looks of it, you used the default setting, but applied it to images with very small faces. This won't work too well.
Please reprocess the images with the small faces again using the optimized for small faces setting, and the IMatch results will be a lot better.
It's a trade-off between recognition rate and performance.

I used a library of 80,000 images containing one or more persons, from all kinds of sources, cameras, motives and image quality. Images spanning 50 years and often 20, 30 or more years of lifetime for each person.
This gives a pretty good sample set with a vast variety of conditions and situations.

IMatch's face AI is really good, because of that.

Are the photos you have posted in your attachment in the open source and have all persons given their permission to post their images on the web?
If not, of you don't know, please remove the attachment. Posting images of identifiable persons without their explicit permission is illegal in many countries, including the entire EU.

IMatch does not use OpenCV for face detection.

All settings you can use to control how IMatch performs face recognition are of course documented. I suggest you refer to the documentation: Face Recognition Options (https://www.photools.com/help/imatch/#people.htm?dl=h-121) This will also explain why IMatch did not detect many of the small faces, and how you can improve that easily.

QuoteBUT: Automization is much much better in IMatch AND XnView MP has some serious bugs, especially XMP-metadata-writing is not working correctly.

Ah...

Quote
If not, of you don't know, please remove the attachment.

Don't know how to do that.

I have removed the attachment.
Please don't post photos for which you have no publishing rights and permissions.
This server is hosted in Germany and EU laws apply.

I pay license fees for the images I use for IMatch documentation and help to be legal.
And if I post another photo, I make sure it is in the PD and I'm allowed to use it.