The Software Engineering department at my FRC team is working on an offseason project. We are in need of multiple images of the following field/game pieces:
- Loading Bay
- Power Port
We looked through the images provided by FIRST on supervisely, but we found that there was very little variation between most images. We would appreciate any images or videos that you can contribute, and we hope that other teams who are in need of said images and videos can find them here.
Have you looked at the field tour videos that were posted right after kickoff?
Hmm, interesting source of video. Thanks for the idea!
If you are trying out Machine Learning, I learned this trick in my Autonomous Vehicle class
You can take a video, separate each frame (lets say 1s video @ 30fps)
There you have 30 images
Then you can go ahead and reverse them so they’re backwards, which gives you 60 images
Then you can do some other things like distort the color, shear, etc. to get even more images. Doing these things can also lead to more robust model as well
Give it a try and see if it works out
Will do! Sounds like a great idea!
We have a ton of images with annotations if you’d like. Around 2000-2500 images. Here’s the link: Supervisely - Dashboard
That is awesome! Thank you so much!
Seeing some of the additional discussion in this thread, I think I have a better idea now of what you might be looking for.
We had a GoPro camera installed on our robot at our first competition very close to where our targeting system camera was located and we filmed a bunch of our matches. For other matches we had the GoPro mounted looking out over the front of the robot near where our driving camera is located. We collected a lot of match footage of our robot collecting and shooting.
If these “robot POV” videos would be helpful, I can put you in touch with our Media Lead who can figure out how to get the videos to you.
That would be great! It would be awesome to use some in-game footage to test on.
FRC Team 5172 got a bunch of POV shots from their robot at an event this year: FRC Team 5172 Gators - YouTube
Keep in mind the networks work best when trained on images from the same camera that is to be used on the robot. Most of our images came from a LifeCam 3000. A model trained on GoPro feeds might not work with another camera. If you want the model we trained for comparison, I’ll be able to send you some stuff within a week.
Yeah, we’re just beta testing right now to see if this is a viable replacement to regular computer vision. Once (if?) we can get back into our school and setup a field, we will take frames from the robot’s onboard cameras and use that.
We’d appreciate having your model too.
Fair enough. Just take the results with a grain of salt if you’re not using the same camera
Will do on model once we finish it.
This link isn’t working for me. Is it correct?
In 2019 we took pictures of Hatch Panels with our phones + used WPILib’s images. The camera on the robot was a Intel Realsense and the model worked amazing.
This topic was automatically closed 365 days after the last reply. New replies are no longer allowed.