In need of high resolution images of vision target

Does anyone have {1080p, 1440p, 4k} images of the vision tape with illumination (either green or something else)? I want to see how long it takes my vision pipeline to process them on the jetson, so that I am able to choose an appropriate webcam instead of buying 3 and returning 2 of them.

Try looking here: https://github.com/wpilibsuite/allwpilib/releases/download/v2020.1.2/2020SampleVisionImages.zip.

1 Like

Those images aren’t even 1080p sadly.

It’s not true 4k, but someone took a picture with the flash on, so it’s good enough for now.

You could always scale up those photos. They will be blurry but the number of pixels will be higher.

1 Like

Do the webcams you’re considering not have the ability to change resolution?

Have you considered how much resolution you need for the accuracy you desire from the distance you’ll be operating from?

Generally speaking, high resolution like 1080 or more is unnecessary and slows down your ability to process a decent number of frames per second. Additionally, if you’re using a network camera or sending the video to the driver station, that resolution turns into a bandwidth hog as well.

In the past, our team has used 720x480 with great success and even less in some cases. Processing at least 10fps or 15fps is desirable to allow for decent response in your feedback loops.

bob - Team 1967 - The Janksters

1 Like

Even that resolution is consider high. Remember that you are limited to 4 Mbps over the field WiFi network. We (and I believe most teams) work with images around 320x240 pixels or smaller. For us, going with a small image and significant JPEG compression means we can run at 30fps, which is much better for image lag.

2 Likes

Thanks to shared memory so no uploading / downloading to GPU is required, and the cuda implementation of find contours, the whole pipeline runs on the gpu. This means that we can get ~40 pose solutions a second with extremely precise accuracy.

1 Like