@veg Thanks! We are getting similar performance using last year’s wpilibpi version on an rPi 4b with a corral. Getting between 14-16 fps. We are labeling and training with Roboflow to TF(TF Record) then converting to TFLite.
We did test a bit with Teachable Machine, but it appeared it will only create classification models, is that correct?
That is my understanding as well.
I don’t want to dilute the LimeLight training effort, but if anyone would be open to collaborating on a set of tagged training images, let’s do it!
We have had success with this workflow for detection with bounding box.
Upload and Label in Roboflow.
Build Dataset on Roboflow.
Export to TF Record.
Convert to TF Lite. Axon 2022 or TF Lite in Google Collab(How to Train a Custom TensorFlow Lite Object Detection Model).
Upload Model to wpilibpi.
Using the basic inference python script from last year you can easily test the model on your device and see the labeled bounding box detection in action.
I was wondering if labelling could be partially automated using a tool like this, or maybe using a basic blob detection routine. But then I think, if an algorithm could reliably auto-label the images, there would be no need to train a net.
Still…
Here is our dataset with 300 labeled images.
We will be adding about 100 more cone images early next week to help improve detection of non-upright cones.
Automated labeling is available. In reference to @joseppi’s post, you can use trained projects on Roboflow for labeling. Whether its your own project, or another one that is publicly available on Roboflow Universe:
By the way, for full disclosure, I work in Developer Experience at Roboflow.
slightly derailing, but is y’all’s dataset public? looked around a bit and couldn’t find anything
ty!
I’m seeing something unexpected here. I am writing an app to run various .tflite models, and coded it to run either with or without a Coral attached. On an RPi 400 with 4 meg of RAM I am getting about the same performance with or without the Coral ~ 7.5 - 15 fps.
I am not running the Coral at 2X clock speed, so it probably can still beat the Pi. But this makes me wonder if teams without a Coral could still think about running ML models.
Have you benchmarked without a Coral, and if so, do you see anything similar?
Yes, we get 1-3 FPS without the Corral. So much lower performance.
have you got the network tables working yet?
Our team has not been using a ML approach this season, so I have not carried this project any further. It should be trivial to take the information that is being imprinted onto the video stream and publish it to the network tables, but I don’t have a working setup at the moment to be able to test this here.
ok, thanks
This topic was automatically closed 365 days after the last reply. New replies are no longer allowed.