Hi CD
Now that FIRST announced the use of AprilTags for robot vision in the field, I am aware that higher performance will be needed to process the images and not lose FPS.
So I’m trying to figure out on how to install PhotonVision on a Jetson Nano.
I hope someone can help me, thank you.
They have pretty great documentation on their website
BUT, they don’t recommend nor officially support a Jetson Nano.
Yes, the two primary challenges would be:
A) Building the native libraries for Jetson
B) leveraging the hardware to its fullest capiabilities.
I’m not sure to what extent A is done - I’ve not worked on a Jetson before and am not sure if the linux builds that exist within photonvision will work. Specifically, the apriltag, cscore, wpilib math, and network tables dependencies will be the big things to be looking into.
I’m sure that B has not been broached yet. PhotonVision does not have support for leveraging hardware acceleration on a Jetson.
For alternatives… April Tags — ISAAC 2020.1NX documentation may be of interest?
WPILib is already built for aarch64 targets, so in theory it should “just work”. I haven’t touched a Nano in a long time, so YMMV.
The stable version install script for Debian systems (assuming that you’re using the jetson-nano-jp461-sd-card-image.zip Ubuntu 18.04 image) works almost out of the box. You’ll need to apt-get install curl first.
USB camera work out of the box. HW support does not (expected) and CSI / MIPI cameras don’t work out of the box. The later would need integration with the Jetson libraries since this part of the pipeline is completely different on Jetson as compared with RPi.
I’ve got some more notes in a doc that I can prep to share if wanted, though we’re probably not pursuing Jetson Nano much more for this right now.
(Oops this was meant to be a general reply.)
All your findings line up with my last experience on a Jetson. Of note is the Nano and TK1 / TX1/2 boards will be losing software support from Nvidia soon, and aren’t worth pursuing further efforts on IMO.
A Pi4 will run circles around a Nano unless the Nano’s GPU is being used.
The Orin Nano will be a good starter platform for more powerful vision than Pi platforms once it is available. I still plan on adding hardware support for Jetson platforms in the future but simply didn’t get time to do so this year.
I realize it may be too expensive for most but I am curious if something like the Arduino Portenta would run photon and furthermore if it would work with the vision shield. Another product would be the Arduino Nicla vision although it would need a Ethernet port.
It will not. It’s just not designed for it unfortunately.
In general, I wouldn’t expect PhotonVision to run on anything that can’t run a proper operating system such as Linux. This means no Arduino, no Pixy, no Pi Pico, etc. You’ll want a SBC, not a microcontroller.
I had success get it working with an ELP USB camera on a Nano by following the instructions for installing on Other Debian-Based Co-Processor. I had to install CURL first.
sudo apt update
sudo apt install curl
As noted above, a PiCam didn’t work, only USB cameras.
As far as performance, it seemed to be on par with a LimeLight 2 for AprilTag detection.
I would have guessed Nano to be 10 to 100 times faster judging from all the Internet blogs.
This topic was automatically closed 365 days after the last reply. New replies are no longer allowed.