Object Detector using Python 3

Note: This assumes you have set up the NVidia Jetson as is described in the jetson repository README.

If you want to try out TensorFlow with Python 3 rather than Python 2, you'll probably want to set up everything in a separate workspace and overlay that on the previous one so you don't mess up all the Python 2 packages.

Setup virtual environment for Python 3:

pip3 install --user virtualenvwrapper
export WORKON_HOME=~/Envs
export VIRTUALENVWRAPPER_PYTHON=/usr/bin/python3
source ~/.local/bin/virtualenvwrapper.sh
echo "export WORKON_HOME=~/Envs" >> ~/.bashrc
echo "export VIRTUALENVWRAPPER_PYTHON=/usr/bin/python3" >> ~/.bashrc
echo "source ~/.local/bin/virtualenvwrapper.sh" >> ~/.bashrc
echo 'export PATH="$PATH:$HOME/.local/bin"' >> ~/.bashrc

mkvirtualenv -p python3 --system-site-packages tf-python3
workon tf-python3
git clone https://github.com/jetsonhacks/installTensorFlowJetsonTX.git
cd installTensorFlowJetsonTX/TX2/
pip install tensorflow-1.3.0-cp35-cp35m-linux_aarch64.whl
pip install catkin_pkg rospkg pillow
deactivate

If using Python 3.5, as described you need cuDNNv7. Download the .deb from https://developer.nvidia.com/nvidia-tensorrt3rc-download. Then:

sudo dpkg -i nv-tensorrt-repo-ubuntu1604-rc-cuda8.0-trt3.0-20170922_3.0.0-1_arm64.deb
sudo apt update
sudo apt install tensorrt python3-dev

Make ROS work with Python 3:

sudo apt-get install python3-empy # Errors building messages without this

Download this repository.

mkdir -p ~/catkin_py3/src/
cd ~/catkin_py3/src/
git clone https://github.com/WSU-RAS/jetson-py3.git ras_jetson_py3

Clone the vision_opencv package to make cv_bridge work with Python 3:

git clone https://github.com/WSU-RAS/vision_opencv.git

Now build OpenCV 2 for Python 3 (src). Note you'll need roughly 12 GiB of disk space for this. If you don't have that, then rsync a bunch of your files off the Jetson, delete them, and then copy them back after you do the make install. Also, make sure you run sudo ~/jetson_clocks.sh before doing this so it'll compile faster.

sudo apt-get install libavcodec-dev libavformat-dev libswscale-dev libv4l-dev \
    libxvidcore-dev libx264-dev libgtk-3-dev libatlas-base-dev gfortran python3-dev
wget https://github.com/opencv/opencv/archive/3.3.1.zip -O opencv.zip
wget https://github.com/opencv/opencv_contrib/archive/3.3.1.zip -O opencv_contrib.zip
unzip opencv.zip
unzip opencv_contrib.zip
cd ~/opencv-3.3.1/
mkdir build
cd build
cmake -D CMAKE_BUILD_TYPE=RELEASE \
    -D CMAKE_INSTALL_PREFIX=/usr/local \
    -D INSTALL_PYTHON_EXAMPLES=ON \
    -D INSTALL_C_EXAMPLES=OFF \
    -D OPENCV_EXTRA_MODULES_PATH=~/opencv_contrib-3.3.1/modules \
    -D PYTHON_EXECUTABLE=/usr/bin/python3 \
    -D BUILD_EXAMPLES=ON ..
make -j6
sudo make install
sudo ldconfig

Now, if you do "import cv2" it'll still try to use the Python 2.7 version provided in ROS, which will error. Thus, first try the version we just installed. Before you use it, you need to:

export PYTHONPATH="/usr/local/lib/python3.5/dist-packages:$PYTHONPATH"

which we will do in setup-env.sh before running the Python 3 object detection.

Make sure you have sourced the previous workspace before building, to make this an overlay workspace:

source /opt/ros/lunar/setup.bash
source ~/catkin_ws/devel/setup.bash

Build everything:

cd ~/catkin_py3
workon tf-python3
catkin_make -DFILTER=OFF -DPYTHON_EXECUTABLE=$(which python) -DPYTHON_VERSION=3

Finally, run TensorFlow with Python 3:

cd ~/catkin_py3
. src/ras_jetson_py3/setup-env.sh
roslaunch ras_jetson_py3 object_detector.launch

Or, run components individually:

roscore
rosrun image_view image_view image:=/camera/rgb/image_raw
rostopic echo /object_detector

cd ~/catkin_py3
. src/ras_jetson_py3/setup-env.sh
rosrun ras_jetson_py3 object_detector.py \
    _graph_path:=~/networks/ssd_mobilenet_v1.pb \
    _labels_path:=~/networks/tf_label_map.pbtxt