Training an object detection model - ML on Raspberry Pi with MediaPipe Series

Рет қаралды 6,539

Күн бұрын

Пікірлер: 26

@GoogleDevelopers 10 ай бұрын

Resources: Getting started with object detection→ goo.gle/41pWKCv MediaPipe → goo.gle/3u7sbos Object detection model customization guide → goo.gle/3stlY5N Hyperparameters → goo.gle/3QxZR6k Demo → goo.gle/3SvLKAS Label Studio→ goo.gle/476VqXg

@santiagolassog 9 ай бұрын

Hello Google Dev Team! This new way of training with Mediapipe is excellent, especially for deploying on Raspberry Pi 4. I have trained a custom network to identify industrial vehicles within a factory, and it works ✅. I would like to improve the processing speed because sometimes it is a bit slow. I would like to know if you can do a video explaining how to train a network for a tflite model to use Coral AI on Raspberry pi for improve a speed, and do the corresponding deploy 😁. I thank you very much, congrats! Excellent work.

@oliverexcellent 9 ай бұрын

Thanks for the great video. The tips on how to fiddle with Label Studio's labels.json file output (make sure you have 0 be a background class, etc.) were extremely helpful. (FWIW, my current project is a Pi + a battery-powered water gun. It's meant to detect and deter squirrels trying to eat oranges on our orange tree.) I think a great addition to the series would be some intro tips on how to adjust hyperparameters to potentially improve accuracy. Please consider putting it on the to-do list!

@paultr88 9 ай бұрын

Good call! I need to learn more about that myself, but it's definitely something I'll look into :)

@thomashu1095 6 ай бұрын

Hi Paul, thanks for your clear instructions. I was able to create a simple model, deployed to Pi and ran it successfully. I also managed to compile it to a Edgetpu model using Kang’s colab by down grade to tensorflow 2.13.0. It did compiled. Unfortunately, it failed to run on Pi for some runtime errors. Would you please provide instructions on how to compile it with mediapipe support. Thanks!

@connormark1727 Ай бұрын

Is there any way to determine the number of nodes and layers(pooled and convolutional) in the CNN that MediaPipe created? Thank you

@hasankesoglu8 7 ай бұрын

Hello, your video is really useful, thank you. There is one point I don't understand. I see that you created a label for the background. Are there background images in your dataset and have you labeled them? Or do we just need to create it as a label and automatically the unlabeled area in any image becomes the background?

@ptruiz_google 6 ай бұрын

Any unlabeled area takes on that label :) It's a generic catch-all

@嘆甲郎 10 ай бұрын

I have trained a model now. Can I add new photos and categories for training based on this model? If so, will previously trained parts be retrained?

@paultr88 9 ай бұрын

Nope unfortunately once you start that training process it'll remove the images it currently knows and replace them with whatever the entire set is that you're using for the new training.

@frankbraker 4 ай бұрын

at 12:39, there are some other commands "qat_hparams = object_detector..QATHParams(..." that never got executed in this tutorial? What's that about? Thanks for this awesome guide too! That was very helpful. I'm working on a calibration fixture, and ideally I would like to import this custom model into OpenCV.

@paultr88 4 ай бұрын

Quantization! Basically reducing the model size, but that gets more into the data science side of things. It's worth running, but I didn't want to make the video even longer by getting into it :)