Face recognition woefully off mark

DeepakP

Please find an image of my screen where the subject is shown as the picture where the app suggests this picture is included but as you can see it is way off, any suggestions on what is going on?

Andrey

DeepakP If you use search on this forum, you will be able to find multiple discussins on the same topic.

Long story short, face recongiton is not a magic, it is just a program. And it is not perfect and will never be, at least until we remove the requirement that it should work on any regular PC, not require $25K+ GPU and do it reasonably fast. And as it is with any machine learhing algorithm, “garbage in - garbage out” rule applies. That means if you only have handful of low quality images with obscured faces, or not looking into the camera confirmed for that particular person, you should not expect that that will be enough for program to understand who that person is and come up with high quality suggestions only.

In order to improve quality of the suggetsions, you should confirm enough of high quality faces (hight resolution, good lighting conditions, not obscured, facing straight to the camera), and once you do that, accuracy of suggestions will improve.

Also, if you make mistakes and accept wrong suggetsion for that person, that also drives the accuracy down a lot. It makes sense to manually verify already accepted faces and remove incorrect people.

Also, if it happens that for some reason majority of the faces you have for some person are actually taken from the side or even from the back, and this is just what it is, you can ask the program to limit the amount of suggestions using “Suggest less” feature in the context menu of the person.

Hope that helps.

Tony

DeepakP but as you can see it is way off

Hi Deepak. It’s part of the training of the algorithm. The human brain is amazing that we can identify a person even by a side profile or even by someone’s posture. The software doesn’t have our neural abilities. The best thing is continue training TonFotos and eventually it will settle down. So on the photos in your screen shot, select them and mark them as “Ignore this face suggestion” and it won’t be used again (Shift-backspace or right click and “Ignore this face suggestion”.

I know the feeling is to use every photo but focus on using the non-fuzzy photos and the ones with as much of the face visible as possible. Resist the urge to use every photo in facial recognition.

If the photo is important but the face is not very recognizable, rename the photo and add the person’s name. Facial recognition won’t find it but you can search for it.

DeepakP

Thank you for your detailed response and I do understand the “art” of facial recognition. What surprised me was how far it was from the primary image, and only on this one person. For all others it was doing what I have to expect from this amazing package, hence my question. But will persevere.

Tony

Facial recognition seems off for the person for whom I have the most photos (hundreds to thousands). TonFotos is generally very accurate where I have few photos of a person in which case it is close to 100% accurate. It could be where I have thousands of photos of the same person, then the facial recognition must deal with many more variations of that person. I end up with the same results as you, other people, hands, side views of other people, rocks, patterns on walls, etc.

Evonet

Whilst appreciating very well that facial recognition is still more of an art than a science, I notice two things I find disappointing - perhaps you can improve these easily?

Firstly, having identified the only person in a photograph and given them a name, I still find the same photograph cropping up as suggestions for other people. I imagine this is a fairly simple fix in software?

Secondly, I do a lot of work in Africa, and about half the people in my photos are black and half white. The program often suggests lots of black matches for someone I have defined as white-skinned, and vice versa. This does seem a rather obvious mismatch to catch.

Having said all that, Tonfotos is the only program I’ve found that does exactly what I need!

Andrey

Evonet Firstly, having identified the only person in a photograph and given them a name, I still find the same photograph cropping up as suggestions for other people. I imagine this is a fairly simple fix in software?

I’m afraid not… Application can only measure similarity between faces, it has no idea if this is the same person or not. However, if this face pops up as a suggestion to other peple, that means it actually looks close to some other faces already confirmed for those people. Otherwise it would not suggest it. This can be typical for low quality faces (like on the image in the first post here - blurred, not facing to the camera, etc). Those faces typically look more or less the same for the program, that is why there are so many issues with them.

Evonet Having said all that, Tonfotos is the only program I’ve found that does exactly what I need!

Thank you for kind words! I am glad it is good match for you.

Evonet

I have also experienced the following:

I give a name to someone in a slightly fuzzy photo - and I then find I have to spend an hour or more, deleting totally ridiculous suggestions, a hundred or so at a time - even after selecting “Offer less suggestions”.

As I have several not-that-sharp photos, I’ve now spent more than a day doing this. Is there anything that can be done about this?

Andrey

Evonet Well, the general suggestion is to start working on the person from high quality images facing to the camera. When there is enough of them to form a base understanding who the person is for the program, then you can go marking lower quality faces for that person. However, when you start from lower quality faces, the result is typically what you have described. It is hard to blame the program though, there is not much data it can pull from those images anyway. Even we mostly recognize people on those photos based on context (cloth, body, the event), not so much face itself, while program only looks on the face.