ML photo rejection feels a little mean!

Please note: I am not looking for opinions from regular forum members on this topic.

Last night I took new images for an allotment PoI, and a replacement of the same image for the sign was immediately accepted, but a new picture showing some of the actual allotment was rejected. The images are attached below.

Why aren’t different images allowed into community voting nowadays? A UK reviewer would have no trouble recognising this as a photo of an allotment. I have previously added an allotment image to a different PoI, and it was allowed into community voting, so what changed and why is this image now unacceptable?

I do not have appeals to waste on photos. I have a months long backlog of rejected submissions that need appealing before I can start using them on things like this. I would really like to understand the issue here as to why ML had to reject vs just simply allow into normal voting?

1 Like

Hi

So that people can understand and respect your request at the start, are you only seeking a comment from the wayfarer team?

Otherwise no critique of the actual photo?

Or are you open to a discussion about the variability of the aut assessment system for photos, whilst not commenting on this instance?

1 Like

Yeah general discussion of ML photo rejections is fine, if someone else has experienced similar things examples are cool

Just do not need lots of people commenting their opinions on my photography skills and suggesting I did something wrong.

Noted that you are using yours as a starting point for discussion, so I hope people respect the request.

1 Like

I don’t think you did anything wrong. And let me say upfront that i do not have an official information about how ML works. I am basing this on my own experience with it and what i’ve observed from others.

I don’t think ML is smart enough to know what the sign is for. A human brain says “allotment = garden, therefore a photo of a garden is good.” But the ML can’t make that logic leap. I think it does a one to one comparison of each new photo to the original and compares how similar the subjects are. “Sign vs sign = approved” “Sign vs garden = substantially different, rejected”

Now, why it jumps to rejected instead of passing it to the community, IDK. That’s some secret in the way this ML works that i don’t expect them to explain to us.

Yeah I wasn’t expecting the ML to auto-accept, since its a very different image, but I was expecting it to go to community review like what happened the last time I did this kind of photo addition.

I am worried if this is going to happen more often as it will just promote ongoing signfarer where adding any new, more creative images, or changing from a photo of a sign to the actual object will no longer be feasible without appeals.

1 Like

I would like an explanation of how we can ever get a new kind of image accepted ever again tbh. Assuming we cannot use appeals, since appeals are needed for so many other eligible actual nominations rather than edits

I don’t think we can ever get that explanation. Niantic will never tell us how to bypass the automated photo (and text, not relevant here) filter.

The photo is fine and should have gone to reviewers, but I can see why the automated system struggled. I also don’t think it would be easy to relax the automation to allow this photo through without the unintended consequences being too large.

It would be easier to appeal this if an appeal got returned immediately if successful.

I wouldn’t assume that a single rejection means we can’t ever get anything new past ML. I’d want to see a larger pattern of rejections like this before coming to that conclusion. (I work with data and process improvement, so i always fall back on patterns and trends.)

For a while there, we could appeal rejected edits to help chat. It would be great if we could return to this instead of having to use one of the formal appeals on them.

Edited to include the link about that:

1 Like

Something must have changed, because a few months ago this went to community review. So the ML also has an example-locally- of a similar picture being allowed into voting and accepted for an allotment with a similar sign too being the main image. But it has learnt from this (presumably) and decided to stop me repeating this kind of edit :sob: hence why I think there might end up being a trend towards only very like for like replacements ie sign for sign, building for building.

That would be awesome

I guess I should also link a topic about appeals to help chat ending so I don’t confuse anyone. Here is one:

That is a major leap logically and severely flawed. You have presented one example and cannot jump to such a strong conclusion.

I can see significant differences between that allotment photo and the one that was automatically rejected. Drawing deductions from them being treated differently is of little value.

Had to do some digging…:joy:

Here are an assortment of photo edits that I have had over past 9 months, different locations, and I think a distinct difference from the one already in place.

So I think it’s not about it being a different view, and I thankfully I don’t think there is evidence of a general trend.

My observations from the inception of Emily has been that It struggles when shapes and colours blend too much - it can’t see the wood from the trees very easily, or if something has the appearance of ordinary. No inside knowledge, observations from experience that could be well off the mark. I have no idea how much difference there might be if any between photos and whole submissions :woman_shrugging:

All I know is that when Emily rejects it feels very frustrating as it isn’t just about the reject but that I will probably have to use an appeal. That is not a great user experience.

2 Likes

I agree with @salixsorbus that these 2 allotment photos are significantly different. I have refrained from commenting on the quality of the allotment photo in this case because you specifically said you didn’t want that feedback. But if you’re going to present 2 garden photos, we will have to discuss how they differ.

I did have it happen to me but in the opposite for a new nomination for an Allotment.

Originally sent with a photo of the actual allotments, Rejected by eMiLy.

Sent it with a picture of some metal gates, tiny sign with the allotments initials and a notice board with a couple of A4 pages that where too faded to read, sent to Reviewers who Accepted…

I had an allotment rejected before the automated checks. It was of the allotments, not the gate. Looked like allotments, with lots of greenhouses showing, but reviewers didn’t like it for being allotments.

So considering the original discussion topic. Which I hope I am paraphrasing correctly

It is around the auto system and photos?

I guess a question could be if photos can be auto accepted - which is absolutely great. :tada:

What should happen to the rest? There are 2 piles auto reject and send to community.

I would like to keep the auto system in place because the accepts do encourage me to update photos reflecting changes and more than anything quality of resolution.

Does that mean I then have to accept the fate that it might go into one of those other piles. Roll the dice for the community outcome ( personally I don’t like how it is set up in voting)

Or accept that Emily has rejected…….and what next does it need to be an appeal on the same level of any other rejection or could as @cyndiepooh suggested, their be another route for human review.