Co processor Question

[techhelpbb]

Originally Posted by adciv

Woops, I missed that it had an internal battery.

If anyone has some vision benchmarks between ODroid, Kangaroo, and Tegra TK1, I'd be interested. As an aside, it is legal to use a 2nd RIO as a coprocessor (albeit expensively, R11)

Marshall posted some performance work on the Tegra in the past.
There are multiple ODroid boards so which ones interest you?
Which vision system on the Kangaroo?

I can't necessarily do this myself right now.
These are the questions one would need to know.

[adciv]

I'll take whatever info is out there, this is for general background knowledge on their comparative capabilities. I've wanted to run vision co-processors in the past. The only reasons we aren't this year is the roboRIO is 'good enough' for what we need and our development resources are too low to spend time on the tegra.

[techhelpbb]

Originally Posted by adciv

I'll take whatever info is out there, this is for general background knowledge on their comparative capabilities. I've wanted to run vision co-processors in the past. The only reasons we aren't this year is the roboRIO is 'good enough' for what we need and our development resources are too low to spend time on the tegra.

Only thing I have handy for the ODroid XU3 (last version - the XU4 is even faster):
http://www.eecs.berkeley.edu/Pubs/Te...S-2014-117.pdf

I have a Kangaroo on my desk but I don't really want to put the NI stuff on it.
Maybe I'll order another...Done
Microsoft loves me considering I just renewed my MSDN license...
ODroid loves me because I now own 4 of these XU4 and all sorts of gizmos for them.
Terasic where is my account e-mail?!

[marshall]

Originally Posted by techhelpbb

Marshall posted some performance work on the Tegra in the past.
There are multiple ODroid boards so which ones interest you?
Which vision system on the Kangaroo?

I can't necessarily do this myself right now.
These are the questions one would need to know.

Our work on the Tegra Systems has been interesting in terms of performance. Some folks last year (and this year!) have complained that our code performance was slow on the Jetson. We were running a cascade classifier last year that ran around 15 frames per second on the Jetson TK1. That's not slow in our experience but hey, what do we know?

Running the same code on the TX1 is substantially faster. Moving from a cascade classifier to neural networks is proving to be tricky because we don't have enough training data but also significantly faster too. It also helps that we are training with a Titan X and a 12 core Xeon... shame we can't put it on the robot...

Was our code last year super efficient and optimized for Big O or whatever nonsense you want to measure performance with? No. Ain't nobody got time for that.

Ultimately though, it's all about iterating through code revisions and optimizing the whole package, both the code and the processor.

[techhelpbb]

Originally Posted by marshall

Was our code last year super efficient and optimized for Big O or whatever nonsense you want to measure performance with? No. Ain't nobody got time for that.

Ultimately though, it's all about iterating through code revisions and optimizing the whole package, both the code and the processor.

Agreed and one of the things that makes me think the students I mentor might not field the XU4 is that they only asked for it the 1st week of the competition.

Success with this is very much based on what you put into it.
Sure they have what they need from me right now but it is already week 2.
6 weeks might not be enough commit for them to optimize either.

It doesn't bother me. If this rolls into the summer it's still a good use case and there's enough hardware I can work with them till next year and open source the results.

If I were a team hurting for resources and unsure of the skill level we have I wouldn't be rushing to implement this right now. Might be better to wait for some more HowTo and example cases.

Yikes: the cheapest shipping option for the Terasic board is $25? The board itself is only $99.
Seriously?:

[Kind Warning]

Please note that Terasic only has a Taiwanese Citibank account (BENEFICIARY BANK Name: CITIBANK TAIWAN LTD., SWIFT CODE : CITITWTX) with beneficiary name TERASIC INC., if you receive any instruction asking you to wire payment to another account, please stop the payment and contact us immediately.

We would also like to remind that Terasic Sales Team only uses the domain name at terasic.com or terasic.com.tw (e.g. sales@terasic.com, NOT sales.terasic@dr.com), please pay attention to the domain name when contacting our Sales Team member.

Glad I used a proper credit card not a debit card for that

[ThomasClark]

Originally Posted by adciv

If anyone has some vision benchmarks between ODroid, Kangaroo, and Tegra TK1, I'd be interested.

I don't have benchmarks, but the Kangaroo doing basic target tracking in GRIP seems to get around 10-20 fps