Face recognition terrible

Paco Cabrera

Honestly, although I really like the app, the facial recognition system is terrible! I spend hours recognising faces that are perfectly recognisable and having to confirm them over and over again!

For example, in the Apple Photos program, once a person is recognised, all other photos and new photos that are added of that person are automatically tagged. I think it’s good that I ask you to confirm those that are doubtful or not distinguished well, but having to be again and again recognising perfectly recognisable faces does not seem “very intelligent” the AI that organises the photos.

I hope that one day this problem will be solved, but until then I refuse to waste hours and hours recognising faces that have already been recognised before and that are perfectly recognisable.

I put an example capture… It seems to me that this person is perfectly recognisable, and it is not normal that I get dozens of photos of her to recognise them continuously! And if you add that to the fat that I have thousands of photos in my system, it becomes desperate!!! The app should be to make things easier, not to make you waste hours and hours with this topic.

Andrey

Paco Cabrera Sorry you feel this way.

Unfortunately, every facial recognition technology is not perfect. It makes mistakes that is why user input is important. Though it might seem Tonfotos requires more user effort for processing suggestions than other applications, but the fact is, using Tonfotos you can achieve much better results that with any other photo management app with facial recognition I know of. By result I mean that most faces of people you care about are marked and with minimal errors. Apple Photos is not even close, actually it is one of the worst among software of similar type. And that is because they chose automated approach. Yes it is almost fully automated, but as a result I have multiple duplicate persons on my MacBook and coverage is very low.

We decided to go slightly more labor intensive way, but it produces much better results.

However, we also introduced a lot for tools to make it less labor intensive, but still require confirmations from users. Did you know that you can confirm multiple faces at once, or do you still confirm every face separately? I would encourage your to check our YouTube channel and watch our videos, I bet you will learn a lot about efficient ways to work with faces.

I don’t want to go in detailed discussion why we chose this approach. That is a long story and I don’t think it will convince you anyway. In the earlier versions of Tonfotos we actually had automatic confirmation for very similar faces, but we decided to get rid of it later. So no, we are not coming back to automatic confirmations, since overall it does more harm than good. Sorry if that is not what you expected.

Paco Cabrera

Sinceramente no estoy de acuerdo con lo que me dice. Como puede ser, que la IA que utiliza Tonfotos, pueda dudar entre ejemplos como las imágenes que adjunto? Cada vez que abro la app, tengo que perder horas rectificando las dudas que me propone y realmente esto no es práctico. Con fotos de Apple esto no ocurre en absoluto, una vez identificada una cara, todas las que lógicamente pertenecen a la misma persona ya las cataloga automáticamente, y solo hay que rectificar las “dudas razonables”. Buscaba una alternativa que me me permitiera catalogar mi colección de imágenes guardadas en discos externos y en principio Tonfotos me pareció que sería una buena alternativa, pero visto lo visto ya estoy cansado de perder horas y horas y lo voy a dejar.
Adjunto algunas imágenes de muestra… Como puede la IA dudar con imágenes como las que adjunto como ejemplo?

Andrey

Paco Cabrera Well at least the first example does not look that obvious as you might think. That totally can be the same person, just in different age. Also, I think you are too optimistic to expect 100% facial recogniton accuracy for photos when persons face is hardly visible, like in the very first photo. Even human being will not be able to confidently recognize people in such conditions.

As to other two mathces, it looks as obvious mistake, but if you compare these two faces only. However program provides suggestions based on ALL confirmed faces of this person, not just the one you shown here. And I am sure there are some very low quality faces that you have confirmed in the past for that person. and this suggestion is matching those faces, not the one you displayed here. So one of the most typical reasons for wrong suggestions is mistakes you made by confirming wrong or very low quality faces in the past for that person. I would suggest you to verify just in case faces that you have already confirmed in the past for that person.

As program digs thru faces in your library, it starts with suggesting most confident matches, but after some time there is no more good matches to suggest, but it will continue looking and looking. And when you almost finished mathcing all faces in your library, all it has to suggest are those wierd matches, as it does not have anything better to suggest. This is actually not a bad thing. This means you are mostly done with marking up faces. And if you are bothered with wrong suggestions, you can always use “suggest less” function, and that will stop.

To sum this up, if you are not satisfied with suggestions accuracy, there are two things to consider:

Verify that all confirmed faces for that person are actually correct. Wrongly confirmed faces can attract more wrong suggetsions.
You can use “suggest less” option for that person to limit suggestions to very confident ones.

Stig

I woudl quess and expect that you have discussed the AI of face recon and how to use it the best way..
But people have different needs..
Some like to have only few faces of family, and thats it. While others have large colllections with family, classes, fotball-team etc..
This is why I would like an advanced options.

1.) To set face recon to 99% at first… then a first run..then 98% second run..etc.. This would also help you getting an indication back at what treshold the current model is able to get god results by reporting back to you here.
2.) Be able to capture all faces and put everyone of them into their own face ID. A huge workload scanning all of course.. but just leave the job running for a week(s), and be able to browse all faces and give them names accordingly.
3.) I’ve seen so many differnet “faces”, of clothes, hands, wall… I feel it would be much better to actually train the system with activating one folder at the time.. so I would get more accurate result - (I know which faces are on the pictures in that folder).

Again, I’m happy with the software and hope this might help see things from another angle/person 🙂