Still Images wrongly tagged with language when there is no text

Thank you very much Grayson!

So sorry for the late reply, is there a way to avoid this incorrect tag by the API? Please excuse my ignorance and thanks again!

Hi Grayson.

I hope this new API is under regular review for accuracy and relevance.

As a regular updater of image tags, I too have noticed a steep increase in the application of dubious language tags.

I have narrowed the problem down to three categories of problem; two of which are definitely API specific and the last - possibly exacerbating the API problem - is an IMDb issue/policy...

1) Language spotted but detected incorrectly. As can be seen from the following images:

There are four words visible; one a proper name, one an English word and the remainder (both the same) Estonian. No Spanish at all.

As with the preceding image, this has an Estonian 'Rescue' or 'Salvation' vehicle. The word Pääste does not appear in the German lexicon.

2) API seeing words that are not there. For instance:

No Italian here... no anything here.

Nothing English in this image. What partial letters there are hardly qualify as words.

Really? Exp?

3) This one is more important, harder to quantify and more impactful, but is also the one IMDb will defend; Irrelevant text. Take a look at the following:

Good shout. 'Edge' is definitely English...

...as is the phrase 'Life Jacket'.

'EST', however, is - in this case at least - an abbreviation for the country code 'Estonia' rather than the Latin 'it is' or 'it exists'. Finally...

'BA' may be a half decent qualification to have, but the abbreviation (wherever it may be found) is not a word. 'Latte' is a word but first, it is French and second, it is on a 'watermarked' Instagram tag.

The point is, the API is finding words everywhere.

I assume that the new language tag feature is a prelude for, after enough images have been tagged, the implementation of another filter, such as the ones that currently exist to enable narrowing images down to just those that are red carpet events, those that show a specific actress or just the theatrical posters.

Such a feature would certainly be useful to those users looking for DVD boxart or posters in their native language and, I reluctantly concede, may even be interested in movie screenshots that feature specific languages.

I would dare to suggest, however, that the latter filter use would be carried out infrequently and with a view to 'significance' rather than 'existence'. For instance, referring to ALL 9 images above, I cannot conceive of a useful purpose for the words seen, especially when their usage is incidental, the language has been mis-identified, they turn out to be abbreviations or acronyms or are just plain non-existent. (And all those examples were taken from the images of just one movie.)

I would propose that, as with so many things in life, size matters. The bigger the visible words, the more relevance or usefulness they are likely to have. The following examples make my point;

I suggest that if a couple of modifications were made to the API code, you could avoid >75% of language tag errors:

1) If an image is tagged as 'Poster' or 'Product', allow full detection and language tagging. Most posters have clear, recognisable text and will likely be the most cross-referenced for language.

2) If an image is tagged 'Event', disable language detection altogether. Most event shots have actors stood in front of 'Comic Con', 'Vogue', 'Hyatt Regency' or similar backdrops. I imagine this would be pointless tagging and the least cross-referenced for language.

3) If an image is tagged 'Still Frame' or 'Behind the Scenes', apply a modified detection algorithm that looks for clear, identifiable (say, with >85% certainty) whole words that are, at the very least, 32 pixels high. At an average 96dpi, this would be approximately 8mm. At higher resolutions, this would be smaller, but higher resolution images also tend to be larger, which would compensate for the reduced clarity. This size check would mitigate the identification of most, if not all, of the first nine images' useless or outright incorrectly tagged words.

I suppose this would depend on IMDb's tagging intent. Do they just want to identify every incidence of every word in every image, regardless of the utility of the resultant (and no doubt humongous) database? Bear in mind that if so, there is a resultant, storage and processing cost associated with the identification and later, when implemented, filtering of such an index.

Or, as I fervently hope, do they want to identify useful incidences of useful words that might appeal to fans, researchers etc, with a much reduced overhead? Your move IMDb.