Get Started         Best Practices         Developer Guide         FAQ        

Preparing reference images for Image Recognition

Last Updated: Apr 30, 2018

In order to make your items (objects) recognizable you need to represent them using reference images. You can use one or several reference images to represent an item. Good reference images are crucial to boosting the image recognition performance and keeping your users happy.

The view of an Image Recognition item in the CraftAR Service (see screenshot below) shows an Image Quality field with an orientative 5-stars rating indicating how well that specific image is expected to work in normal conditions when used as a reference image for IR.

Test your reference images

Note that the Image Quality rating we provide does not take the repetitiveness of the pattern into account, which may also negatively affect the recognition performance.

Hence, the rating feature mainly serves to identify reference images that will surely not work well. So it is always wise to eventually test the recognition on the actual image or object, too, to make sure that the experience works the way it is expected.

Our Image Recognition technology relies on the texture (visual elements) in the images to recognise them despite having different viewpoints, changes in illumination, occlusion, etc.. This is why the highest quality scores are given to images that show rich texture.

The remaining of this article will help you to learn how to prepare reference images for your use cases. Note that most of the tips described below are specific to reference images, and you do not need to apply them to query images (read the separate article about those).

This article is divided into the following sections:

Objects that work well

Important: All objects with a certain texture or markings work well with our technology, e.g.:
  • CD/DVD and book covers
  • Newspapers and magazines
  • Logos and brands
  • Posters
  • Packaged goods
  • Monuments and places
In general, the more uniquely textured the object’s surface the more reliable and resilient it is to reflections, occlusions, tilting, and blurring during recognition.



Figure 1. Examples of perfect reference images for objects that work well with our technology:
  1. textual logos with large letters;
  2. unique graphic designs;
  3. labels for a packaged product;
  4. paintings;
  5. textured roller-ups;
  6. drink cans;
  7. logo on a hoodies;
  8. photos in a print article;
  9. facades of Restaurant ( © User:Rdikeman / Wikimedia Commons/CC-BY-SA-3.0).
 

Objects that do not work well

Although we are working hard on improving our technology, certain types of objects do not work well with our solution.
Important: Objects that do not work well include objects that are pattern-less (e.g. with only solid color), are dynamic in nature (e.g. animals), or have thin bodies (e.g. cables, pencils, etc.).



Figure 2. Examples of reference images depicting objects that do not work well: 
  1. solid color rectangles and frames forming part of multiple real objects (i.e. visually ambiguous);
  2. almost completely texture-less white wall
  3. texture-less transparent plastic glass;
  4. pen with narrow body, i.e. its textured surface is very narrow;
  5. texture-less black bag;
  6. solid color cable, i.e. elastic and thin object;
  7. small dense text – but note that textual logos with large letters, e.g. Figure 1.a work very well;
  8. common repetitive texture;
  9. animal, i.e. dynamic in nature.

Note that some of the above examples (e.g.: (h)) would obtain high Image Quality scores because they have strong visual pattern well distributed all over the image.

Picture composition

Important: A good reference image should only show the object that you want to represent. Every item (object) should be represented by at least one image depicting its frontal face.
Other important things to remember are:
  • tilted poses should be avoided, unless you are intentionally using several tilted images;
  • blurred images affect the recognition performance;
  • irrelevant backgrounds around the object do not affect the recognition as long as they do not cover more than 20% of the area of the picture; take into account that the stronger and more complex the irrelevant pattern around the object the greater it will affect the recognition results;
  • in general it is advisable that reference images predominantly show the portion of the object containing its distinctive pattern.
The example below shows how to prepare the best reference image for recognizing this robot design on a business card.



Figure 3. Selecting perfect reference image for recognizing this robot design on a business card:
  1. query image example where the logo should be correctly recognized;
  2. perfect ref. image to accomplish such recognition; it depicts sharp un-tilted version of the logo;
  3. sub-optimal ref. image with wide irrelevant background;
  4. poor ref. image showing complex irrelevant background;
  5. poor ref. image showing tilted logo;
  6. poor ref. image showing blurred logo.
 

Number of images needed for one item

Important: In most cases one reference image is enough. Even in cases of non-planar objects, such as cans or bottles, using one reference image showing their most representative frontal face gives good results.
However there are some special cases where additional images are helpful:
  • Objects that have different faces, e.g. book covers, cereal boxes, etc., often require one image for each face to be recognized.
  • Items with several appearances, e.g. a company logo or a packed product that evolved during the years need one image for each distinctive appearance.
  • Extra tolerance to viewing angles may be obtained by including additional reference images showing the item viewed from the most likely angles — see example demonstrating recognition of event exhibition stands.

Couldn't find what you were looking for?

support@catchoom.com
https://cdn.desk.com/
false
catchoom
Loading
seconds ago
a minute ago
minutes ago
an hour ago
hours ago
a day ago
days ago
about
false
Invalid characters found
/customer/en/portal/articles/autocomplete