Hello! I’ve been doing annotations for a few days now, and I have some questions that I couldn’t find an answer to on the wiki about how I should do certain tasks.
First: when a single label like EU Organic may have multiple ways to annotate it, for example you could just label it as EU Organic, or specify “IT-BIO-009” and similar, or annotate as “Argiculture EU/non EU” which is sometimes written under it and as I understood counts as a part of it. Which of those ways to annotate a certain image is preferred, and whether I should reannotate existing EU Organic labels to a more precise way to mark them (the IT-BIO-009 example)
Second: any time I try to find something with Cyrillic script in the infinite annotations I get irrelevant results in other languages (for example a brand written in Ukrainian language just yields me other random brands in Greek and arabic irrelevant to my query). How do I go around this and whether it’s a bug?
Third: It is said on the wiki that when annotating images, it is mandatory that the image annotated does not contain two different logos at the same time, or it should not be annotated. Does it apply only to regular annotation, or to “Labels & Claims” tab as well, and whether it applies when I am answering regular questions instead of annotations?
Fourth: when answering regular questions, sometimes a label that the app asks about is present on the packaging, however the image of the label detected by the AI in the right lower corner is incorrect. Do I still say that the label is present on the packaging (because it is) or do I say that it isn’t? Or am I supposed to just skip such questions?
Fifth: I couldn’t find how to properly annotate nutrition labels and would like some advices on that
Sixth: the Hunger Games page misses a link to F-Droid for the app download (in the page’s basement) and Mastodon social for OFF organization. I think those should be added (yeah not a question but a suggestion but still)
For EU Organic, you should indeed annotate them as en:eu-organic, as it’s the class associated with the logo. This annotation is then used to predict a category for new logos, so we should choose a logo class that is not too narrow (as it will impact negatively the prediction accuracy). Regarding the organic certification organization ID (such as IT-BIO-009), we already extract those automatically using OCR in Robotoff.
Can you give me an example (by sharing the URL of the search for example)?
What’s the “Labels & Claims” tab you’re refering to? Can you provide me a link? Overall, this rule does not apply for questions, as we’re not using the answer to the question to annotate the logo (when the question was generated from a logo prediction)
If the label is present on the packaging, you should reply yes, even if the part of the image the information was extracted is not true.
Are you still talking about logos? What are nutrition labels?
For the fifth question, there is “Nutrition Label” type of a logo in annotation tasks, including infinite annotation. I know not what it is supposed to mean afar from the information I found on DuckDuckGo
Seventh question arose: In QR Code annotation I see a lot of Data Matrix logos that were annotated as a QR Code. Even though Data Matrix is a type of barcode, I couldn’t find a single source on the internet listing it as QR Code. I suppose that means that all those annotations with Data Matrix as QR Code are incorrect and should be re-annotated to “Not a logo”, however I still want to know if I’m right about this to not mess up anything accidentally.
Though, all of those labels just use “Type: Label” for annotation. Why is there “Nutritional Label” logo type specifically for them, especially considering I haven’t seen anyone annotate Nutriscore with “Type: Nutrition label”. I also can’t find any documentation on that, and it seems most logos that were annotated with “Type: Nutrition Label” are just regular nutrition labels like: “1400 kJ on 100g”, “25g of sugar per portion” and similar stuff
Am I supposed to annotate anything with “Type: Nutrition Label” at all then?
When thinking about adding a new type, we must think about the use we’re going to have of this new logo type. Here, for nutrition labels that can be found on front packaging, we already detect those using the nutrition table detection model, but so far we had no use of them.
So I believe you can safely annotate those with type “other”.
I tested the fix, and it seems to be able to find the brand from the issue (люкс) properly in infinite annotation, however when I tested other brands written in cyrillic (non-Ukrainian language though) it yielded zero results (even though I’m sure there must be results since I was annotating those brands myself), examples include: ярмарка, савушкин, любятово.
That also makes me wonder, that if Data Matrix is actually supposed to be annotated as QR Code, are other types of barcodes supposed to be annotated under type “qr_code” too?
I’m sorry if this is mean to ask this twice, but since I haven’t received a response on this in more then a month I really feel like I need to, to resolve this confusion around which logos fall under “qr_code”.
I know this may be a dumb 8th question, but:
In Data Annotation guidelines, it is said that if more then one logo is “visible” on the photo, it should not be annotated. Just how visible the second logo needs to be, in such case? Like, would a few pixels on the photo edge of another logo being visible count? Or does it also need to be recognizable?
Hello, sorry for the late answer!
There is no specific guidelines currently for QR codes. This type was added in case we wanted to detect QR codes in a future pipeline, as the base of a future detection model.
In my opinion, you can include Data Matrix symbols.