MoveNet on TensorFlow Lite for Microcontrollers

General Discussion
, ,
#1

I am planning to deploy the MoveNet pose estimation model on microcontrollers like the ESP32.

But I am not sure whether the MoveNet model architecture is supported by the TensorFlow Lite for Microcontrollers.

Can I get some advice on it? Thanks.

#3

Hi, is there anyone that can provide me some advice?

#4

@petewarden who might be able to help

#6

It looks like the standard int8 version of MoveNet on TFHub is about 2.7MB in size:
https://tfhub.dev/google/lite-model/movenet/singlepose/lightning/tflite/int8/4

This is probably too large for most embedded boards to store. I would also guess that the latency would be very high, since it’s designed for mobile phone CPUs. I would look into whether you can use a Raspberry Pi instead, if possible.

#7

Thanks for the information.

I think Raspberry Pi would be able to support the MoveNet pose estimation as the tutorial and example were provided:
examples/lite/examples/pose_estimation/raspberry_pi at master · tensorflow/examples · GitHub

#9

Yes just tried verifying an int8 version, on Ardunio 1.8.19 int ESP32 Cam got the following, BUT may be able to load a model some other way.

text section exceeds available space in board. Sketch uses 3507390 bytes (111%) of program storage space. Maximum is 3145728 bytes.

Global variables use 101168 bytes (30%) of dynamic memory, leaving 226512 bytes for local variables. Maximum is 327680 bytes.
Sketch too big; see https://support.arduino.cc/hc/en-us/articles/360013825179 for tips on reducing it.
Error compiling for board ESP32 Wrover Module.