Make object detection faster by using Coral

Learn how to make your object detection model run faster using Google Coral Edge TPU in this final episode of Machine Learning for Raspberry Pi.
00:00 Introduction
00:15 What is EdgeTPU?
01:24 Connecting Google Coral USB Accelerator
02:11 3 steps to run an object detection model on EdgeTPU
02:36 Step 1: Compile the TFLite model
04:15 Step 2: Install the EdgeTPU runtime
05:20 Step 3: Enable EdgeTPU when running the model
07:27 Demo
08:43 EdgeTPU model code
09:14 Coral’s repository of pretrained models
09:36 Thank you for watching!
Colab notebook demonstrating how to compile a TensorFlow Lite model for Edge TPU → goo.gle/3G4RxUH
Instructions to install the EdgeTPU runtime → goo.gle/3xY3oBb
Sample app to run TensorFlow Lite object detection on Raspberry Pi → goo.gle/3GaABw3
Coral examples → goo.gle/3Dup355
Watch all Machine Learning for Raspberry Pi videos → goo.gle/ML-raspberrypi
Subscribe to TensorFlow → goo.gle/TensorFlow
#TensorFlow #MachineLearning #ML #RaspberryPi #EdgeAI
product: TensorFlow - TensorFlow Lite, TensorFlow - General; fullname: Khanh LeViet;

[@TensorFlow]

Subscribe to keep up with the latest in TensorFlow!

[@seyedalihejazi4926]

can you make a video about speeding up inference with raspberry pi video core IV module? I want a cheaper solution to speed up my neural net.

[@miki31nmh]

Muchas gracias amigo, tengo ya hace un año el USB Coral acelerador y los 4 videos es la mejor enseñanza que he recibido. Saludos desde Perú.

[@basavarajgangadhar7521]

Hi Thanks for the video. I see videos only with Raspberry pi and not with windows. I tried working on windows and facing the issue of "coral. adapter" is required. is there any solution for this.?

[@biskero]

does the NVMe module with 2 Edge TPU can be used together or separately for 2 cameras?

[@LucianoHAlbano]

Hi! Excellent your videos! How do I perform some action using Raspberry's GPIO ports after finding the object in image processing? For example, opening a door or turning on a lamp.

[@nikshimah691]

Hello sir, thank you so much for this! However, may I know if I have an existing code which is for a line detection and I would like to just make it run faster, can I use coral accelerator for this purpose?

[@johnalderson7893]

Hi! Does someone know why I am getting almost the same FPS with the TPU? Using the model "efficientdet_lite0.tflite" (not using TPU), I get 6.3 fps, but when I try to use the compiled model "efficientdet_lite0_edgetpu.tflite" (using TPU) I get 7.6 fps. I will highly appreciate if you can help me.

[@PraiseTheLord527]

Hello John, I truly hope that you were able to solve your issue. But were you able to compile the efficient del architecture (not the custom android pig) used in the first video to an Edge TPU model? I would be very grateful if you could show me guidance if you were able to do so. Thank you

[@tim_pearson]

Please can you get together with the Pi foundation to sort out the problems with using the PCIe or M.2 versions on the Pi CM4 - it is crazy that the top AI chip for edge work will not work with the top SBC. Even though it won't run much faster at all it will allow us to run the dual board, and it will mean we can package it into a product.

[@bibibiti1098]

Great job keep going 🌻👍🏻☺️ great work 😘

[@mikolajwojnicki2169]

The Coral USB has been sold out everywhere I've searched! When/where from can we expect more of them?

[@Taeivas]

I believe it has a backorder time of 42 weeks currently on mouser, expected delivery time would be 28th of February currently. Was also looking for those in all the places I knew, not much luck, also don't want to fork out 200€ for one on ebay.

[@DevinFlake]

It can take a while. I ordered one last year around this time and it didn't show up for months! But now that I've got it I'm excited for this project!

[@Taeivas]

@@DevinFlake Ordered both the Dual TPU and USB Accelerator, apparently the Dual TPU had a shorter lead-time, but has barely any support in terms of usage outside embedded applications and motherboards, especially due to it being dual channel. On a positive note, after tinkering to offset the performance cost I managed to scale down a UNet model to run fast enough on CPU to be viable on the edge, was expecting to use one of those to be viable.

[@fabricedao3158]

Thank for the serie !

[@MIKtechnical]

Sir can produce voice the detected object.

[@pondskoi2073]

Can we use it with Laptop

[@Hulkmb]

ERROR: Cannot install tflite-model-maker==0.1.2, tflite-model-maker==0.2.0, tflite-model-maker==0.2.1, tflite-model-maker==0.2.2, tflite-model-maker==0.2.3, tflite-model-maker==0.2.4, tflite-model-maker==0.2.5, tflite-model-maker==0.3.3, tflite-model-maker==0.3.4, tflite-model-maker==0.4.0, tflite-model-maker==0.4.1, tflite-model-maker==0.4.2 and tflite-model-maker==0.4.3 because these package versions have conflicting dependencies.

[@elvenkim]

I have this error when running step "from tflite_support import metadata5'. 'MetadataPopulator' object has no attribute 'load_metadata_and_associated_files' . Anyone can help?

[@PraiseTheLord527]

Hey Elven, Were you able to find a solution for this? I'm facing the same issue...

[@elvenkim]

@@PraiseTheLord527 sorry I am stuck there also

[@mohammedameer9909]

Create a video on using neural compute stick from Intel

[@qzorn4440]

wow This should run much faster on an Orange Pi-5? 😎 Thank you.

[@frankoptis]

Only 3 times faster wouldn't be worth the money. I hope this is only a glitch. I bought the edgetpu and need more than 10 times speedup. I'm using the edgetpu with J4105 Intel CPU.

[@granatapfel6661]

Bring not axis aligned bounding boxes to object detection 😉

[@techsavy5669]

Nice video. Thanks for sharing. Tech companies have made things are so simple, will that be the Road to Insanity of the future Or Road to a new heaven, time will tell! Cheers.

[@user-ms8ei9le7x]

TF the nice ⚡

[@GeneWildhart]

Can't buy a Coral EdgeTPU anywhere. They are all sold out.

[@louisviciedo]

Hi do you have an Coral Dev Board install on a Intel NUC 11?

[@TensorFlow]

Intel NUC 11 is a computer by itself so you can plugin a Coral USB device to its USB 3.0 port and run model inference on Coral EdgeTPU.