After almost all faces are tagged, lots of additional faces get added

dimjon

After almost all faces are tagged, lots of additional faces get added automatically. The newly added faces have no similarity to that person at all. It seems like Tonfotos just dumps a lot of faces to a person that it can not automatically assign, just to put them somewhere, causing a lot of work. Anyone else has this problem?

Andrey

dimjon Situations like that are not so much a bug, rathen than situations 🙂 The typical reason for that is that there are some low quality faces among ones you already confirmed, or maybe even wrong ones. It makes sense to go over already confirmed faces and and clear obvious mistakes. This can improve the situation, but will not guarantee that there will be no more “garbage” suggestions, since there can still remain correct though low quality faces, that are the rason for this noise. In that case you can use tell Tonfotos to offer less suggestions for that person (by right click on the person name in the list on the left). It will increase the threshold for suggestions, only ones that are very similar will be offered, “more or lesss similar” will not be offered any longer.

I would also recomend reading this topic about recognition accuracy and its natural limits: https://community.tonfotos.com/d/228-people-tagging-logic-seems-to-have-gone-astray

dimjon

Thanks a lot for the explanation. Would it be possible / make sense to add a feature like “Filter by low match quality”?
The reason is: My Photo library is HUGE! So it would be really helpful to have a filter that shows you only those already confirmed face matches that are somehow outliers compared to most of the other confirmed faces, compared to go over thousands of photos manually.

Andrey

dimjon This idea popped up few times before, obviously. However, if you will identify those low quality faces, what would you do? What if those are still correct faces? There can be many partially masked faces, looking out of the camera, captured in the dark environment and with low resolution. However, it seems that people would still like to have those marked up. The goal of AI is to help us with this markup, not the vice versa.

dimjon

It would still save a lot of work, don’t you think? As you wrote previously, it makes sense to go over all already confirmed faces to weed out the obvious mistakes. In some cases, the already confirmed faces might hit a 4-digit number. If I go through all of them, it’s much harder to spot them, and quite error-prone. Having the face-selection already filtered down to the obvious outliers, it’s much more convenient and quicker (especially when combined with the image preview I mentioned in the feature request section, as most outliers from my experience appear in situations where lots of unknown people are in the pictures, such as conferences or school assemblies 😉 )

Andrey

dimjon Thank you, that makes sense. However, the thing is I made quick experiments and found out that first of all, some error will not end up being outliers, and that is saddest part. And among outliers wast majority will be just low quality faces. That is why I am not very confident this tool will be very useful. At lest, if you do it simple way, maybe some additional research is necessary to find better algorithm.

Steffi

Thanks a lot @Andrey for the great explanations in this thread and the one you referred to.

To add on this I would have two questions/suggestions:

First, once I selected the “offer less suggestions” option, is there a “way back”? So, is it possible to decrease the threshold later, for instance, if I get too few suggestions after the threshold was increased? It would be really cool to have something like a slider, which has a default value, but I can increase or decrease the threshold stepwise for each person. I see that this might be too complicated for “standard users”, but how about adding an “advanced mode” so people that want to have more influence on the face recognition process have additional options?

And another idea related to the “how to handle low quality photos” topic:
I have the same problem that the more low quality photos I add for a person, the worse the new face suggestions become. For example, if the algorithm finds some pictures from COVID times, where the peorson wears a face mask, I know that the next suggestions will become worse once I confirm this picture. But still, I want to have this person added to the photo. Wouldn’t it be an idea to be able to confirm the person but without adding it to the training data? Like two options of confirming the photo, with and without adding it to the training sample? Is see that finding these outliers automatically as suggested before is hard, but if I could do this manually, this would help.

Thanks and best regards 🙂

Edit: Sorry, I just saw that the button changes from “offer less suggestions” to “offer more suggestions” once pressed. Should have tried it first 🙈 A slider would still be cool to be able to play around with the threshold.

Andrey

Steffi First, once I selected the “offer less suggestions” option, is there a “way back”?

Sure, it is right there, in the same menu.

Steffi Is see that finding these outliers automatically as suggested before is hard, but if I could do this manually, this would help.

We definitely will be working furhter to find a solution to that issue, and, this particular idea is being around for a while. This is one of the ways to aproach it. However, the preference is to find a solution, that does not requeire user to be an expert in machine learing. I can’t say what will it be, but we will defeintely be woring in this direction. Right now, unfortunally, we are deailing with few more pressing issues. But this one is defeintely in the roadmap, this way or another.