Query images – new size limits

Recently, we have updated the limits in our API which we set for query images. We think we should describe these limits in more detail, to guarantee they are properly understood.

There are three attributes we use to filter out query images: file size, image dimensions and image resolution. While the first two are pretty straightforward, the latter needs more explanation. Why? Well, it is important to understand that, what directly influences the accuracy of recognition result, is the resolution of an object-to-recognize, not the resolution of the image containing that object. To put it in more illustrative form, having an object-to-recognize of resolution i.e. 50x50px, you will get no results regardless the resolution of an image, be it 50x50px, 100x100px or 5000x5000px.

We impose these limits to provide our algorithm with input which yields unambiguous results. The number of objects and the area occupied by them is unknown upon receiving query, thus we cannot verify whether each object-to-recognize exceeds the limits. We impose the limits on the whole image instead.

Each mode, i.e. single or multi, has its own restrictions:
– minimal image dimension is 100px
– minimal image resolution is 0.05Mpix (i.e. 260×195)
– maximal image resolution is 0.31Mpix (i.e. 640×480)
– maximal file size is 0.5MB

Note that in single, by stating ‘minimal/maximal image resolution’ we also mean minimal/maximal object-to-recognize resolution.

– minimal image dimension is 100px
– minimal image resolution is 0.1Mpix (i.e. 372×279)
– maximal image resolution is 5Mpix (i.e. 2560×1920)
– maximal file size is 3.5MB

Note that, in multi, each of the objects-to-recognize should follow minimal resolution restrictions from single.

Images following the above requirements, but containing objects-to-recognize that do not, will probably return ambiguous recognition results or no results at all.

Please also note that we chose the limits to ensure the best user experience. In our various tests we have noticed, that images satisfying these requirements provide the most concrete and the most accurate results within a reasonable time.

Below are examples of bad and good query images for multi mode. Objects-to-recognize, that follow requirements are outlined in green, otherwise in red. Objects not to be recognized are not outlined.

Bad query image:


 Good query image:

Piotr Gibas,
Recognize.im Developer

WordPress SEO fine-tune by Meta SEO Pack from Poradnik Webmastera