Hi all
How did your team tackle vision this season? Do you think you were successful with your setup? What might you change in the future? Thanks
0 voters
No double vote option? My team did photon vision on a small cluster and then did ML vision processing as well.
we utilized hardware for both the Limelight and Photonvision (BeeLink Mini PC), and both software for Photonvision and Limelight.
it got a bit clunky especially when trying to properly organize and structure it in our code. other than that, we were successful in making it work.
the end goal is resorting to a custom vision solution on some multi-camera setup with multiple coprocessors, and (hopefully) implementing ML.
LL+LL software
What was successful in our set-up was that non-programming that wanted to “dabble” in coding got a chance to make a measurable performance increase to our robot without having to go through multiple weeks of training first.
PV with custom hardware. Dual cameras. One monochrome Global Shutter for Apriltag. The second was a color webcam for retro targets. Not that it matters going forward, do to the move away from retro targets, one change we should have made this year would have been a webcam with a narrower FOV.
We ran photon on an OPi5 using two monochrome arducam global shutter cameras with one on each side of the robot. Our swerve odometry was egregiously bad due to what we think was an issue with our gyro so we relied heavily on photon for auto placement. One gripe with it was that multitag was running on the rio and ate some loop cycles, but I believe they are changing this for the future. It was easy to setup and integrate into code, would recommend exploring this option.
how was your experience with limelight vs photonvision? do you recommend one of the other?
though this varies from team to team. i had a fair experience with both the LL and photonvision, i very much prefer photonvision over the limelight though. we only used our limelight to perform retroreflective tape vis. processing, and since the tape is going to be phased out, my team is sticking with photonvision.
now, i must say that they each have their pros and cons. while photonvision does allow you to work with custom hardware, can be much more affordable, and supports multiple cameras, it is less stable than the LL. but limelight does support apriltag tracking and is pretty easy to set up.
ultimately, taking into account a bunch of factors, i’d say do your own research and decide what’s best for your team.
Could you elaborate on this? Why is a wide FOV bad? how big of an FOV is too big?
When aligning to the retro targets, quite often the camera would identify targets on adjacent grid locations. The FOV of the camera we used was approximately 90 deg. 60 deg would have prevented this issue.
We used two Limelights, both only in auton. One of them we used for AprilTag detection, and it worked pretty well. There’s definitely aspects we want to improve, but I think that’s more with our code than the Limelight itself; it detected the tags very well. The other one was used for basic cone detection (essentially telling it to look for the color yellow). This was barely used, but when we used it in the more controlled environment of auton it worked well for lining up to the cone to intake it.
I haven’t used photonvision, so I can’t comment on it, but we found limelights to work very well for our needs.
We had a double limelight setup in the front and back of our robot. Always used in autonomous, aside from our auto alignment system in teleop which was never really used in competition. The pose estimation worked well when set up properly, however we would like to start using more cameras and more custom vision pipelines, so moving towards our own custom solutions is definitely something we’ll be working on this offseason.