Profiling estimates/Emulator for Smart Cameras in VBAI

[b.ploetzeneder]

NI smartcams + VBAI provide a really neat and fast way to get results in image processing - however, predicting their performance is close to impossible. Information like "DSP coprocessor" is nice, but it's not adding any information that I can use.

I have benchmarked the devices that went through my hands with some standard resolutions+ROIs and algorithms, so I have some idea, but most of the time I just wonder "Would this work on a 1774?" "Would it work on something slower?" I usually call an engineer and try a loan device, but this takes time and rather static concepts.

I know that algorithm runtimes (especially the more complex ones) strongly depend on the sample data, so it's not just a matter of scaling, but even an estimate (like 8-12ms) would improve my experience rapid prototyping.So, I'd just use Profile->Performance Meter -> Profile for NI 1772..
 

[Elmar_NI]

Hi,

I have to say that SC DSP models are not recommended anymore. The reason for that is that the latest 177x generation is much more powerful as the DSP model. Anyhow the only way to get a correct performance value is the way that you already descriped. Even if we test all kinds of resolutions and functions the test is not really comparable.Because the performance really depends on the settings in the algorithm e.g. geometric match (rotated,scaled,occluded).

So you get at least 4 different results (1.rotated, 2.scaled,3.occluded,4. All three parameters are checked).

The speed is also image content related. For example a big barcode with a lot of edges will take longer as a simple black image with a white rectangular.

We did a benchmark to compare the different smart camera models. I hope it gives you an idea how the performance difference looks like.

Elmar

[b.ploetzeneder]

Thanks for the benchmark link, I've not seen that before.

I'm very, very familiar with the fact that timing strongly depends a lot on a number of parameters (I've been playing around a lot with them for some years now), but still, if I use good sample data + parameters and profile my code, I can get something very close to the actual performance on a PC. I'm aware that this doesn't scale in any way linearly to different hardware (mainly memory structure and parallelization), but I still think that there could be some estimate that's not totally random.
(Although, some of your algorithms, especially the barcodes have some really cryptic timing behaviour..)

Anyways, thanks.
The reason I was asking for this was because while I love your cams, I would require faster models with low resolution. 177x unfortunately are not high-performance enough for me in many cases.
 

[Darren]

Any idea that has not received any kudos within a year after posting will be automatically declined.