Hello,

Apologies for not responding sooner. Key members of our team have been on holiday and I'm still awaiting a response on ImageTarget detection performance compared to Cylinders for logos.

Regardless, today I tried another experiment using Vuforia's new 6.0 SDK. I focused upon detection performance of the Pepsi Max logo only on the cylinder target:

https://www.dropbox.com/s/yj0x4yz9th7mhqz/cylinder_detect_pepsi_max_logo.mp4?dl=0

I used the same dataset and cylinder target as in my previous reply to this thread. As you can see, detection performance is excellent using the Galaxy S6 Edge. The only time detection performance suffers is in the beginning when the camera is not focused on my mock label. However, after focus is achieved, detection is nearly instantaneous each time.

From this experiment, my anticipation is that the team's opinion be similar to the results: There is little to no detection performance difference between the target types.

Thanks

Hello,

Apologies for the late response, and for the snafu with your forum post.

I will reach out to core engineering to see if my assumption that planar target performance, specifically for logos, is equivalent to cylinder targets is correct. I will respond to this thread once I have more information.

Thanks

Hey,

Thank you for the walk through and the dataset, I was able to take these and recreate similar results as shown in your video. Unfortunately, this is at similar level to where we were before when carrying out our different tests.

The Vuforia detection is working great but only when it sees other parts of the cylinder target, not the Pepsi Max logo. The logo will be the main target which people will be trying to scan. This is shown in your video at 14s where there is no recognition until the cylinder is rotated to present nutritional information. Once the detection has happened Vuforia does an impressive job at tracking.

I tested this by removing all of the other information off the target and creating a new dataset which I have attached below. With this I cannot achieve a detection.

I still can’t seem to get my head around why the cylinder target can’t recognise the Pepsi max logo but the image target can? You mentioned before they are using very similar if not the same technology for detection and tracking.

We would much prefer to use the cylinder target because of the advantages if offers especially when scanning a can.

Thank you
Stuart

Hello,

I was able to take the artwork you provided and make a quick mock up test. As you can see, I achieved excellent tracking performance:

https://www.dropbox.com/s/frehl97ra2tlbnb/pepsi_cylinder.mp4?dl=0

I believe this problem is related to the dimensions you used when creating the cylinder target. Admittedly, this can be a tricky exercise that I had to think through carefully. Here are the steps that I took:

1.     Obtain the dimensions of the CanArtwork_PepsiMAX.JPG artwork (=2500H x 1301W)
2.     Determine the scene unit length you want to specify (for CylinderPepsiMaxCanMeasured.zip, you used 90) - this corresponds to the artwork's height in #1
3.     Use an aspect ratio conversion tool to calculate the new height-to-width ratio, based upon the scene unit length used. This will give the updated "width" in scene units (=173) - this corresponds to the artwork's width in #1. Note that the aspect ratio of the original artwork is preserved, which is critical to detection and tracking performance.
4.     Considering that the scene unit "width" of the can is actually the circumference, convert this to the diameter (=55)
5.     Create your cylinder target using: Length=90, Top Diameter=55, Bottom Diameter=55

You can download the dataset I created here: https://www.dropbox.com/s/10bfnqsp8hawrdu/CylinderPepsiMaxCanMeasured.zip?dl=0

Thanks and good luck!

Hi Medabit,

We are trying to get this to work with a standard UK 330ml Pepsi Max can, in regards to dimensions we have tried two different approaches both relate only to the cylindrical bit not the complex tapers at the top or bottom of the can.

Approach 1 – Take the dimensions from the client supplied artwork, these were a little big compared to the sample so I reduced it down by a factor of 50 to be fore like the sample sodacan.xml. Diameter: 56.14, Length: 91.8

Approach 2 – Take measurements of the can and adapt the artwork to be the right ratio. Diameter: 65, Length: 90

Both approaches do work eventually but with all the problems mentioned in previous posts. I have attached three zip files, for both of these approaches and the flat image targets for comparison.

Thanks,
Stuey

Hello,

I'm assuming that this is a standard Pepsi can? When you entered the dimensions for the cylinder target, did measure the top and bottom diameters at the lip of the can? Or from the widest point where it starts to taper in towards the lip? Using the former could have an effect on detection/tracking performance as we do not support "complex" cylinder shapes such as a standard soda can.

Finally, please also provide me with the cylinder device database (zip file containing .xml and .dat files) downloaded from the portal. These should be the same that uses the artwork you've attached here.

Thanks.

Hi Medabit,

Thanks for the reply, I’m glad you can see the confusion I am having when it comes to changing between image targets and cylinder targets. The difference in performance is like night and day and I can’t really figure out why.

In answer to some of your questions, we have tested both on a real can and on printed art work wrapped around a can. Both result in pretty poor detection and tracking especially when compared with the flat image target which seems to work almost instantly.

When doing our investigations into cylinder targets we also came to the conclusion that the environment was having too much of an effect on the can with shine and reflections obscuring critical markers. Since testing with the flat image target the environment does not pose an issue with the shiny can, I am even able to cover some of the markers with my thumb and it will still detect and track with great accuracy.

I have included both Image target and cylinder target in my posts below, do you think It could be to do with the way I am setting these up within the target manager? Or is there something else which would explain the differences between the two tracking methods.

Stuey

Hello,

Thanks for the feedback, and apologies for the delay in response.

Upon a strictly visual assessment of your logo, I too believe that this should provide a suitable level of detection and tracking performance as a planar target. It is unclear to me why performance would be worse if this is converted to a cylinder targets as I was under the assumption that the underlying tech was the same.

When testing with the cylinder target, are you using a real can for testing? I'm wondering if specular reflections, coming from environmental lighting, could be adversely affecting performance. In our labs, we create mock can artwork by printing the target and then wrapping it around a real can. This preserves the size and shape, but mostly eliminates light reflections. Have you tried that approach to testing your cylinder target?

Thanks

Update:

So after a large amount of research into how cylinder targets work, what attributes produce the best results and how the environment may effect detection and tracking. I unfortunately came to the conclusion Vuforia cylinder targets wouldn’t be a suitable solution for the project I am working on. As far as I can tell this has nothing to do with Vuforia (which seems unbelievably good) but more to do with the level of detail provided within the image target.

There is not enough detail within the Pepsi Max logo to generate features required for detection and tracking. An example being the globe which is an organic shape, Vuforia only shows four features with no matter how much tweaking can’t be improved. This means the feature detection is relying on bits of text and other elements placed around the can, not the main focal point on the front.

When I was almost about to give up, I decided to test out standard flat image targets. After optimizing the image following the developer best practices, I exported the database and swapped it out in the sample project and the results it gives are fantastic. Almost instant detection, which tracks with great reliability on shiny materials in environments which are less than optimum.

There is still a fair bit of work to get this to a finished state but defiantly showing progress. One of the issues being that because Vuforia is trying to detect a flat image it is constantly try to determine which way is up, causing the tea pot to twitch. I am hoping that I will be able to smooth this out so that it isn’t as noticeable on the final animation.

My question for the community is that how come the image target is working so well compared to the cylinder target which is more or less useless? They are both using the same graphics with the same features?

I would prefer to use the cylinder target as this seems to offer better tracking and the ability to interact with the cylindrical object in 3d space. Anyone willign to offer any suggestions?

Stuey