does mapillary team use the submitted blur requests to train/teach the auto-blur ?
on one hand, it would be terribly disappointing if all the user-submitted blurs would go to waste. that’s a massive amount of information. granted, the requests shouldn’t be taken as “this is the absolute truth about blurring on this image” but more like “this is better than what the image had before”, but that’s detail
on the other hand, i’ve seen people request blurs in images i uploaded that seem to blur “there’s a human head somewhere there” - that is, not faces, but humans from the back, including shoulders and pretty much covering the upper half of the person. that would not be useful for face recognition training at all (and those blur requests got approved, too).
so what is it currently, are the blur requests used for training or discarded ?
They are used for training so keep them coming! And the share of blurring non-faces which you brought an example of is not that frequent so the algorithm is far from thinking that this is what it should do as well.
Thanks a lot to everyone who blurs (and unblurs)! We are planning to give more attention to that in our stats as well since it is really important work and obviously we need our community’s help with it. Thus the huuuuge gratitude we have for you blurrers!
thank you for the info. it does seem that the auto-blurs are getting slightly better, although maybe i’m just imagining that now
if i see somebody request blurring of incorrect objects, should i request a fix on those images then ?
see, you probably have a lot of obsessed perfectionists in the community, so incorrect things kinda irk us
@Richlv, glad to hear it seems like that! I think it is actually true, not just an illusion
Any time that you put into fixing blurs is highly appreciated by us - be it mistakes by our algorithms or mistakes by other users.
It’s part of my job to review blur requests, while when submitting blur requests I am like every other user (so supposed to do it from my own free time). However, also to the perfectionist side, so it has happened a couple of times that when I see faulty blur requests, I have to go and fix them and before I know it I’ve spent a good 20 minutes on it in the middle of the workday.
Another note, sometimes users request blurs that are not really faces or license plates, but they still get approved because we want to respect our community’s sense of privacy and appropriateness (examples: people in swimwear, drying laundry, ads… etc.). Maybe @yubin has something to say what effect this may have on algorithm learning…
The auto-blur definitely becomes better and better, no doubt. Comparing the results I used to have a year ago and those I have now makes it obivous.
Just a question related to your last note above: I am currently uploading something like 4500 pictures of NW China. In a few of of them, a car crash scene is shot. I initially planned not to upload these specific pictures but after having read this post I’ll upload them and blur what needs to. Is this the proper way to go?
I can’t tell right now whether anything wrong/shocking appear on these pictures as it’s difficult to determine on a cellphone screen. On the spot, one could see a badly damaged car and two covered bodies lying on the ground.
Please let me know should anything special (deletion or anything) be done on top of the blurring. I’ll do it and send you a private message once done.
@Shlublu, it’s really great to hear that you notice the improvements. =)
As for the car crash scene, I agree that you’ll get a better overview of what is visible and not once the photos are on a computer screen. And although we don’t forbid these photos, in general it may be best to skip the sensitive ones if possible. So delete, not just blur, because for this specific case the issue is not whether someone is instantly recognisable from the photo, but more the whole picture of it and what is going on, which will probably still be quite clear after blurring.
In addition to this just being too sensitive for some people who may be browsing those photos, there is the extra notion of how would it feel for a person who was personally touched by this crash, i.e. a close one to the victims, to explore around their city or neighbourhood on Mapillary and come across this scene. Even if the chance is small, then the harm can potentially be great which is why I personally would recommend to delete the few photos that document the accident. You have every right to disagree of course.
@Shlublu - good to hear, then at least this time there is no practical problem. And as a general guideline we have instructed (also on our web page) that private things should not be captured, arguably these kinds of situations could fall under that category. But perhaps we should be even more specific, bringing out that sensitive situations should also be avoided…
@katrin Yes, I know the guidelines You should define what falls under this category with examples (the usual “including but not limited to”). Common sense is a good thing but has the drawback of having one definition per person
@katrin, thanks for the updates. regarding blurring of non-faces and non-licence plates, i’d probably be pedantic and allow for the blur request reviewers to specify that some areas are not faces or licence plates, thus excluding them from the training pool - but then i’d probably also require manually classifying/verifying that licence plate and face blurs are marked like that…
guess it all depends on how well the auto-recognition can be trained in all cases and how much @yubin can implement
Until the blurring with Mapillary, I never noticed how littered with annoying false blurs Google Streetview is. Any chance of Mapillary demonstrating how much better they are at it?
The funny and lucky things is that mly is not consequent in the blurrings of text. If I Know that I will need a text, I take three pictures of it. In a text of two words, I may have to combine two pictures to get the words.
Well, to be able to really claim that we should need data to prove it, and while we may be able to get our own “hit rate” then I am not sure Google wants to reveal theirs. Also it may still turn out we were wrong and in the big picture perform worse, even if it may seem at times that it’s better. Bottom line, the most fruitful thing is for us to focus on our own blurring and set the target of constantly “beating ourselves”.
I’m not really sure I can say I have noticed it improving much. I think part of the problem here in the U.S. is that the blur algorithm seems to think all license plates are long and skinny like European license plates. Our license plates have a different aspect ratio but even when one is actually detected by the auto-blur process it is still blurred out with a long bar that sometimes leaves the top/bottom visible while blurring more than needed to the sides. To make matters worse, our street signs tend to be shaped a lot like European license plates so there are a lot of false positives that end up blurring out valuable map data.
Also, it is easily confused if the license plate is not facing the camera straight on. Sequences in parking lots are the worst to manually review.
Interestingly, I just stumbled on a sequence where one license plate in particular was detected with a fairly high degree of accuracy. Start at this picture and move forward in the sequence: https://www.mapillary.com/app/?pKey=lQKeFHeoS0VDLskDKEow-w
The red Mazda had its license plate blurred in most pictures while cars in the right hand lane at a very similar distance did not (I have manually blurred some of them by now so it won’t show up that way if you look now)
They say red cars get pulled over by police more often. Maybe the blur algorithm also prefers red?
Thank you, @toebee. @katrin, I’m with him on the odd shape. I didn’t realize it until I read it. But I see it in a lot of false positives like this one. Once I read @toebee’s comment on it being a different ratio I was like “oh ya, I see that all over the place”.
