Log in or register to post comments

How to improve the accuracy of Object Recognition

October 19, 2017 - 7:41am #1

We are developing a first AR app to show how much sugar is in the food product.

Currently we are at the stage which contains 15 salad dressings. Vuforia can have pretty good accuracy on this category.

However, we've been trying to add 6 yogurts into the app, Vuforia cannot recognize them at all.

 

The yogurt images are in the attachment.

 

Can anyone tell me how to improve the accuracy of yogurt recognition?

Thank you!

 

How to improve the accuracy of Object Recognition

October 19, 2017 - 3:14pm #2

Hello wang39429,

Have you tried using our Cylinder Target feature for these objects? Since they are all pretty simple, cylindrical shapes you might end up with better tracking results.

If this doesn't work for you, or you end up having other objects you need to track, here are some tips for using Object Targets:

There are many factors that can affect the tracking performance of a Object Target, both during the scanning process and when running an app. Most issues with Object Targets can be traced to the creation of the .od file using the Vuforia Object Scanner. Be sure to follow the scanner app instructions: https://library.vuforia.com/articles/Training/Vuforia-Object-Scanner-Users-Guide, paying close attention to step #8 in the article.

  • When creating the .od file, was the model scanned in an environment that was free of background details which may have introduced features that were not part of the model? Scanning in 'cluttered' environments can introduce false detection/tracking points.
  • When creating the .od file, was there any specular reflections on the model introduced by environmental lighting? Scanning objects that have reflective surfaces under direct lighting can introduces areas with no detection/tracking points.
  • Are you using the recommended devices referenced on the tool download page?: https://developer.vuforia.com/downloads/tool

In our labs, we utilize four primary strategies for creating an optimal Object Target scanning environment:

  • All background surfaces are colored at 18% gray. An easy, off-the-shelf solution is to buy bed sheets near this color and drape everything in the environment that could be seen by the scanning device's camera.
  • No direct lighting. We use light boxes and/or diffusers to eliminate direct lighting upon the object and minimize any spectral reflections.
  • Utilize a 360 turntable to re-orient the device. This is especially helpful when you've set your environment to near ideal conditions within a limited area. You can spin the model and scan in 360 degrees without having to move around it.
  • Be sure that the environment in which you’re testing (via the Object Scanner app ‘test mode’) is the same in which you’re verifying tracking (via the sample code). Environmental factors such as lighting, shadows, spectral reflections, etc. can negatively affect tracking performance, so awareness of how the environment is interacting with your model is important for qualifying performance.

Thanks,

-Vuforia Support

Log in or register to post comments