tkDNN-ROS

• YOLO object detection with ROS and TensorRT using tkDNN
  • Currently, only YOLO is supported.
  • (Optional) Image rectification is added - supports pinhole-radtan, pinhole-equidistant (Fisheye)
  • (Optional) Downsampled inference is added (e.g. image is coming in at 30Hz but infer it upto 15Hz)
  • (Optional) Image saving is added - save only detected image in .jpg format in image folder

Comparison of performance and other YOLO implementation? - here

Dependencies

• CUDA, cuDNN, TensorRT (version 8)
• .rt file, built with .weights and .cfg files

[Click to see how to make `.rt` file in detail]

● prepare .rt file ★much work to do★

• Export weight and cfg file into .bin files as original repo

Get the darknet (only for export, not used for detection)
$ git clone https://git.hipert.unimore.it/fgatti/darknet.git
If this darknet repo does not work, try with this one: 
              https://github.com/AlexeyAB/darknet/issues/6116#issuecomment-655483646
              https://github.com/AlexeyAB/darknet/files/4890564/darknet-master.zip
$ cd darknet
$ make
$ mkdir layers debug
$ ./darknet export <path-to-cfg-file> <path-to-weights> layers
-> .bin files are generated in debug and layers folders

• Build .rt file, which can generate executable file

$ cd tkDNN/tests/darknet
$ cp yolo4.cpp <name_you_want>.cpp
$ gedit <name_you_want>.cpp

std::string bin_path = "path from tkDNN/build folder"; //edit

// Edit here with output layer, check exported 'layers' folder
// e.g., for yolo v4 tiny, exported .bin file with name 'g' are 'g30.bin' and 'g37.bin'
// files starting with 'g' are output layer
std::vector<std::string> output_bins = {
    bin_path + "/debug/layer30_out.bin",
    bin_path + "/debug/layer37_out.bin"
};

// also check .cfg and .names (.txt) files directory
std::string cfg_path  = std::string(TKDNN_PATH) + "/tests/darknet/cfg/yolo4tiny.cfg";
std::string name_path = std::string(TKDNN_PATH) + "/tests/darknet/names/coco.names";

$ cd tkdnn/build
$ cmake .. && make

# executable file name with <name_you_want> is generated.
# Excute it to generate .rt file
### it reads the .cfg and .bin files written in <name_you_want>.cpp file, so directories should be accurate
$ ./test_<name_you_want> 

● Execution, refer here for more detail

$ cd tkdnn/build

Edit the paths in the demoConfig.yaml file before!
$ ./demo ../demo/demoConfig.yaml

● Changing inference data type: re-generate .rt file after export tkdnn mode

type one of belows: (TKDNN_MODE=FP32 is default before change)
$ export TKDNN_MODE=FP16
$ export TKDNN_MODE=INT8

and re-generate .rt file as above before execute.

---

How to install

• clone this repo

$ cd ~/<your_workspace>/src
$ git clone --recursive https://github.com/engcang/tkdnn-ros
$ cd tkdnn-ros/tkDNN
$ mkdir build 
$ mkdir installed
$ cd build
$ cmake .. -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=../installed
$ make install

• build ROS package

$ cd ~/<your_workspace>

$ catkin build -DtkDNN_DIR=<absolute_path_to_your_workspace>/src/tkdnn-ros/tkDNN/installed/share/tkDNN/cmake

or

$ echo "export tkdnn_DIR=<absolute_path_to_your_workspace>/src/tkdnn-ros/tkDNN/installed/share/tkDNN/cmake" >> ~/.bashrc
$ . ~/.bashrc
$ catkin build

or

$ catkin config -DtkDNN_DIR=<absolute_path_to_your_workspace>/src/tkdnn-ros/tkDNN/installed/share/tkDNN/cmake
$ catkin build

• Do not forget to register tkDNN libraries into LD_LIBRARY_PATH

$ cd ~/<your_workspace>
$ echo "export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/src/tkdnn-ros/tkDNN/installed/lib" >> ~/.bashrc
$ . ~/.bashrc

How to run

• Make sure you have .rt files with your CUDA, cuDNN, TensorRT version.

• change parameters in main.launch file

• run the code

$ roslaunch tkdnn-ros main.launch