meta-mediapipe

Introduction

This is a hackey copy-paste-smash-together of the meta-tensorflow layer at http://git.yoctoproject.org/cgit/cgit.cgi/meta-tensorflow to instead support building Mediapipe. Tensorflow(-lite at least) is a dependency of Mediapipe but I have made no effort to re-use the tensorflow built by yocto in the mediapipe build.

The tensorflow recipes probably don't even work right now in this layer.

Linked external OpenEmbedded Libraries

For an exhaustive list check out the dependencies listed in the mediapipe bb file.

• protobuf(-native)
• Google projects tend to statically compile protobuf into themselves
• Protobuf statically compiled in exposes all the protobuf symbols
• Using another library which links to or includes protobuf is then impossible
• So we've modified mediapipe to use and link to an externally built protobuf
• We still let bazel download the version of protobuf it wants, to let it use the bazel scripts
• However we modify the bazel files in that protobuf download to expose the yocto-built protobuf
• mesa - required for GPU inference/processing support
• opencv - not really neccessary for most usage of mediapipe
• ffmpeg - largely untested

TODO

• Test and implement building the mediapipe python library
• Achieve compatiblity with the meta-tensorflow layer ** Test and fix the existing tensorflow build along-side the mediapipe build. ** Re-use the tensorflow parts built by yocto in the mediapipe build ** Merge meta-tensorflow and meta-mediapipe
• Use bazel to accumulate more of the library ** Use Bazel to gather only the needed header files ** Use Bazel to gather only the needed tflite files
• Support a more configurable build for the c++ library ** Use bitbake features to enable/disable certain calculators and modules

Overview

Mediapipe is a Media-processing (mainly for ML inference) pipelineing library.

Enabling easy use of hardware acceleration, protobufs, and tflite for feature/object detection and pose inference.

See: https://github.com/google/mediapipe https://github.com/tensorflow

This repository is build configuration for cross-compiling mediapipe using the Open-Embedded/Yocto ecosystem.

See: https://www.yoctoproject.org/ https://github.com/siemens/kas

Usage

I'm not going to go into great detail here but the kas config I used to build a mediapipe c++ library for RPi 4 was:

header:
  version: 10

machine: raspberrypi4-64

distro: our-own-poky-based-distro

repos:
  poky:
    url: https://git.yoctoproject.org/git/poky
    # refspec: master
    refspec: df2a1f37f77d19e62a5fbdb462a9b24e104f0448
    layers:
      meta:
      meta-poky:
      meta-yocto-bsp:
  raspberrypi:
    url: git://git.yoctoproject.org/meta-raspberrypi
    # refspec: master
    refspec: 0e48785e15d864438cade6e526056dbc92e7a690
  private-system-layer:
    # our own layer with OS configuration and networking bits
  private-app-layer:
    # our own software which made use of the c++ library
  oe:
    url: https://github.com/openembedded/meta-openembedded
    # refspec: master
    refspec: e418ee4657e084c8b4d42aabf76ff6df99253e91
    layers:
      meta-oe:
      meta-python:
      meta-multimedia:
      meta-networking:
  tensorflow:
    url: THIS REPOSITORY
    refspec: master

local_conf_header:
  meta-custom: |
    PACKAGE_CLASSES ?= "package_deb"
    LICENSE_FLAGS_WHITELIST += "commercial"
    ENABLE_DWC2_HOST = "1"
    PRSERV_HOST = "localhost:0"
    BB_NUMBER_THREADS = "4"

target:
  # Our own system image targets from `private-system-layer`