We're trying to implement tracking of Merge cube - cube with 6 different textures on the sides, using Wikitude's native SDK. Currently we use ImageTrackerConfiguration with six separate images and compute cube transformation based on the tracked side. With this approach, tracking quality is good when one side is clearly visible, but tracking is easily lost when cube is rotated from one side to another, even by slow rotation.
Are we possibly missing something in the tracker configuration, e.g. some parameters worth tweaking? Or maybe there is a better approach for tracking such markers available in the SDK?
Did you already try to work with our Object Recognition? Your use case sounds similar to what we show in our latest demo for our Expert Edition launch - https://www.youtube.com/watch?v=iak9zBhhoR4&feature=emb_title - and it sounds like Object Recognition might be the suitable approach.
In case you work with Unity, we definitely recommend that you check out also our Expert Edition. Details on the features of this recent launched version of the SDK can be found here: https://www.wikitude.com/blog-augmented-reality-for-experts-introducing-wikitude-sdk-9/. But also our Professional Edition of the SDK offers a comprehensive Object Recognition and Tracking feature.
So give it a try (the technical documentation gives you guidelines on how to create an Object Target) and let us know if you have further questions.
Thank you for your suggestion, tried the proposed approach with object target.
Unfortunately, for now it worked much worse in our case than tracking side images separately. Attached it the best point cloud we were able to get (from ~30 images). The main problem is that it doesn't seem possible to take images from all directions, including bottom side of the cube. And even without bottom side, tracking is established only when looking at the object from nearly the same angles as images were taken.
Could you please also share sample images of the object from the different sides? Then we can check suitability internally.
Thx and greetings
Sure, here are two series of images with different lighting conditions
[Upd] First link was wrong, corrected link: https://drive.google.com/open?id=13AjnLv0ARaX-D-6ieEvQXVzp2NZxYA44
We checked the images internally and our colleagues from the computer vision team mentioned that the object itself should be suitable for object reco. Here are some tips on how to create and improve the object map:
Extending the map with new images is always a good appraoch to enhanve the map and add areas to the existing one; This step-by-step approach allows you to improve the map gradually and test it accordingly. Did you try to extend the map with images of the missing faces?
Please try the above and let us know if the map improved.
Thx and greetings
Trying the proposed approach with building initial map based on 3 adjacent sides and then extending it.
First step works fine, an accurate map with three faces is created.
But when extending it with series of images of another 3 sides (with two overlapping with previous sides, and one new), new side is not added and the map is only slightly rotated.
Tried with three different series of images, but they give the same results: https://drive.google.com/open?id=13AjnLv0ARaX-D-6ieEvQXVzp2NZxYA44 (folders 1, 2, 3 each contain series of photos of three adjacent sides)