Skip to content

Latest commit

 

History

History
173 lines (136 loc) · 5.17 KB

README.md

File metadata and controls

173 lines (136 loc) · 5.17 KB
Raw

Coral Support

Bazel Setup

You can compile MediaPipe with enabled Edge TPU support to run Coral models. Just add --define MEDIAPIPE_EDGE_TPU=<type> to the bazel command:

  • --define MEDIAPIPE_EDGE_TPU=usb for Coral USB devices on Linux and macOS
  • --define MEDIAPIPE_EDGE_TPU=pci for Coral PCIe devices on Linux
  • --define MEDIAPIPE_EDGE_TPU=all for both Coral USB and PCIe devices on Linux

You have to install libusb library in order to compile with USB support:

  • libusb-1.0-0-dev on Linux
  • libusb on macOS via MacPorts or Homebrew

Command to compile face detection Coral example:

bazel build \
  --compilation_mode=opt \
  --define darwinn_portable=1 \
  --define MEDIAPIPE_DISABLE_GPU=1 \
  --define MEDIAPIPE_EDGE_TPU=usb \
  --linkopt=-l:libusb-1.0.so \
  mediapipe/examples/coral:face_detection_tpu build

Cross-compilation

Sometimes you need to cross-compile MediaPipe source code, e.g. get ARM32 or ARM64 binaries on x86 system. Install cross-compilation toolchain on your system or use our preconfigured Docker environment for that:

# For ARM32 (e.g. Raspberry Pi)
make -C mediapipe/examples/coral PLATFORM=armhf docker

# For ARM64 (e.g. Coral Dev Board)
make -C mediapipe/examples/coral PLATFORM=arm64 docker

After running this command you'll get a shell to the Docker environment which has everything ready to start compilation:

# For ARM32 (e.g. Raspberry Pi)
bazel build \
    --crosstool_top=@crosstool//:toolchains \
    --compiler=gcc \
    --cpu=armv7a \
    --define darwinn_portable=1 \
    --define MEDIAPIPE_DISABLE_GPU=1 \
    --define MEDIAPIPE_EDGE_TPU=usb \
    --linkopt=-l:libusb-1.0.so \
    mediapipe/examples/coral:face_detection_tpu build

# For ARM64 (e.g. Coral Dev Board)
bazel build \
    --crosstool_top=@crosstool//:toolchains \
    --compiler=gcc \
    --cpu=aarch64 \
    --define darwinn_portable=1 \
    --define MEDIAPIPE_DISABLE_GPU=1 \
    --define MEDIAPIPE_EDGE_TPU=usb \
    --linkopt=-l:libusb-1.0.so \
    mediapipe/examples/coral:face_detection_tpu build

Our Docker environment defines ${BAZEL_CPU} value, so you can use it directly:

bazel build \
    --crosstool_top=@crosstool//:toolchains \
    --compiler=gcc \
    --cpu=${BAZEL_CPU} \
    --define darwinn_portable=1 \
    --define MEDIAPIPE_DISABLE_GPU=1 \
    --define MEDIAPIPE_EDGE_TPU=usb \
    --linkopt=-l:libusb-1.0.so \
    mediapipe/examples/coral:face_detection_tpu build

The command above is already defined in our Makefile, so you can simply run:

make -C mediapipe/examples/coral \
     BAZEL_TARGET=mediapipe/examples/coral:face_detection_tpu \
     build

The output binary will be automatically copied to out/<platform> directory.

You can also run compilation inside Docker environment as a single command:

make -C mediapipe/examples/coral \
     PLATFORM=armhf \
     DOCKER_COMMAND="make -C mediapipe/examples/coral BAZEL_TARGET=mediapipe/examples/coral:face_detection_tpu build" \
     docker

and get the output binary from out/<platform> directory. Any Mediapipe target can be cross-compiled this way, e.g. try mediapipe/examples/desktop/hand_tracking:hand_tracking_cpu.

To summarize everything:

Arch PLATFORM Output Board
ARM32 PLATFORM=armhf out/armv7a Raspberry Pi
ARM64 PLATFORM=arm64 out/aarch64 Coral Dev Board

Coral Examples

There are two Coral examples in mediapipe/examples/coral directory. Compile them for your platform:

# Face detection
make -C mediapipe/examples/coral \
     PLATFORM=armhf \
     DOCKER_COMMAND="make -C mediapipe/examples/coral BAZEL_TARGET=mediapipe/examples/coral:face_detection_tpu build" \
     docker

# Object detection
make -C mediapipe/examples/coral \
     PLATFORM=armhf \
     DOCKER_COMMAND="make -C mediapipe/examples/coral BAZEL_TARGET=mediapipe/examples/coral:object_detection_tpu build" \
     docker

Copy output binaries along with corresponding auxiliary files to your target system. You can copy the whole mediapipe folder for simplicity:

scp -r mediapipe <user>@<host>:.

OpenCV runtime libraries need to be installed on your target system:

sudo apt-get install -y \
    libopencv-core-dev \
    libopencv-highgui-dev \
    libopencv-calib3d-dev \
    libopencv-features2d-dev \
    libopencv-imgproc-dev \
    libopencv-video-dev

If you are going to connect Coral USB accelerator to your target system then you'll also need libusb library:

sudo apt-get install -y \
   libusb-1.0-0

Connect USB camera and Coral device to your target system and run the copied binaries:

# Face Detection
GLOG_logtostderr=1 ./face_detection_tpu --calculator_graph_config_file \
    mediapipe/examples/coral/graphs/face_detection_desktop_live.pbtxt

# Object Detection
GLOG_logtostderr=1 ./object_detection_tpu --calculator_graph_config_file \
    mediapipe/examples/coral/graphs/object_detection_desktop_live.pbtxt