60 is the recommended limit, but it isn't a hard limit. The effective limit really depends on the total number of features among the targets. Realistically you aren't going to be able to handle 200 targets simultaneously. So you'll need to reduce the search space somehow.

Are there characteristics that differentiate subsets of targets?

Feel free to PM me if you want to describe the implementation privately.

Thanks for the response.

We only want 1 item at a time but out of a pool of 200.
We have a 200 page book and need something different to happen on each page the phone is pointed at. Each page is a full picture and all are different enough that the SDK shouldn't get confused between them.

I haven't tested the limit, but I was under the impression that it was 60 from reading through the documentation.

So you are saying there is no effective limit and is only dependent on how different my trackables are?

Sorry if I'm confused. Thanks again for the help.

There is no limit defined by the SDK, but there is an effective limit due to the cost of data set processing. Eventually detection and tracking performance will decline to an unacceptable level. This is where the recommendation for a max of 60 comes from. It's based on the practical experience of the SDK developers and testers.

If you can repeat a set of sub-images across multiple pages, then you may be able to define a sort of indexing system that would enable you to distribute the target set across multiple datasets.

For examples, let's say that you have a dataset of 10 index images and each index image is associated w/ a dataset of 20 images. When one of the 10 index images is detected, its associated dataset of 20 images is activated. So Index Image 1 is associated w/ pages 1 - 20, Index Image 2 w/ pages 21 - 40 etc..

I have been struggling with this exact issue all weekend!

My issue is similar, where I need more than the recommended 60. I am interested in the indexing system you described. Do you have some example code to show how this would work?

I would still prefer an option where I could have the code cycle through the datasets until it finds a match then load that dataset because in some cases I might not be able to repeat a set of sub images. What do you recommend to achieve this?

Hi,

in version 2 there was introduced option to have more than one dataset active. I couldn't find any mentioning of performance. Version 2 claims to be faster than 1. What are the practical limits in v2 now in total number of simultanously recognized image targets? even split over more datasets.

Thnaks

S.

Hi Stan,

"in version 2 there was introduced option to have more than one dataset active"

...where exactly does it say this?

Cloud Reco is one of the new main features, though at present there is no sample that shows recognising more than one target.

Regarding performance there is a general improvement in detection and tracking, however a lot of this in reality depends upon the device, cpu power, gpu power, camera resolution etc. particularly given that Vuforia works with over 400 devices.  It is difficult to state specific performance across devices, however if there is a problem discovered, we will do our utmost to remedy it.

In theory there is no reason one cannot extend MAX_SIMULTANEOUS to 10 or more, though it will depend on the device.

N

Hi Nalin

here 

https://developer.vuforia.com/resources/dev-guide/transitionmigration-guide

it says

"Multiple device databases can now be activated simultaneously"

and also in API reference and in unity script it is possible to check more datasets as active.

I think you misundertood a bit the topic here. I think it is duscused maximum targets limit in offline databases.

It was 60 I think before in 1.5 and now it sais somewhere that limit is 100 per dataset. But as you can have multiple datasets activated in version 2 in theory there is option to detect as many as you like, but the question is how many in average is wise in terms of performance.

This is not simultanous tracking but detection. tracked can be only one at the same time that's fine.

S.

Hi Stan

Yes you are correct.

I have not used multiple active datasets myself, so I will try to find out more information.

Essentially you want to understand 1) is 100 the limit across all datasets or single dataset and 2) what is the optimal number of images to use per dataset.

Have I understood correctly?

N

NalinS wrote:

Essentially you want to understand 1) is 100 the limit across all datasets or single dataset and 2) what is the optimal number of images to use per dataset.

Have I understood correctly?

Quite close but not exacly. I'm only interested to know what would be a acceptable number of detectable images that device is trying to recognize. It does not matter to me distribution of image tagets across datasets. So no matter whether there are 100 dataset per 10 targets or 10 datasets with 100 targets. Before there was limit to have one dataset active therefore max 60 targets coul be detectable at a time. By v2 this limit is basically gone so in theory you can try to detect plenty of targets using multiple activated datasets but I assume there is some rational limit of what average hardware can try to do to keep user experience at good level. This is what I was trying to get.

That threshold really needs to be determined by testing, because it's sensitive to a range of factors beyonds simply the number of active targets. We've recommended a max of 60 in consideration of the range of devices that Vuforia supports, but there are applications that have exceeded that number. 

Also ancillary processing (e.g. rendering, code execution .. ) impacts the maximum. As you recognize, it's a matter of what processing capacity is available in a given runtime context. 

Thanks David. This is what I thought but because I did not find any guidline like this in v2 documentation and because v2 claims to be faster I thought I'll clarify this. Would be good to cinsider to cut out your answer and include it to docs because v1 was limiting this by API but v2 is not anymore and one can design app thinking he can detect plenty of targets and it may lead to big mistake in concept.