FairMOT-cpp

An inference-only C++ implementation of FairMOT.

Pre-requisites

• CMake 3.18.4
• Python headers 3.8.10
• CUDA 11.4
• cuDNN 8.1.1
• LibTorch 1.12.1
• Torchvision 0.13.1
• OpenCV 4.7.0
• Eigen 3.4.0

Other than Eigen, older versions for the dependencies may also work.

Quick start

1. Clone the repo and its submodules

   git clone https://github.com/liyier90/FairMOT-cpp.git && cd FairMOT-cpp
   git submodule update --init --recursive
   

2. Download fairmot_dla34.pth to the weights/ folder.

3. Convert the PyTorch model to a JIT model.

   cd python
   pip install -r requirements.txt
   python convert_to_jit.py
   

   Older versions of the required Python packages may work as well.

4. Set up third-party C++ dependencies.

   cd ../third_party
   

   1. Download and install Eigen.

      wget https://gitlab.com/libeigen/eigen/-/archive/3.4.0/eigen-3.4.0.tar.gz
      tar -xvf eigen-3.4.0.tar.gz && cd eigen-3.4.0
      mkdir build && cd build
      cmake -DCMAKE_INSTALL_PREFIX=</path/to/third_party/eigen> ..
      make install
      cd ../..
      

   2. Download and install OpenCV

      wget https://github.com/opencv/opencv/archive/refs/tags/4.7.0.tar.gz
      tar -xvf 4.7.0.tar.gz && cd opencv-4.7.0
      mkdir build && cd build
      cmake -DCMAKE_INSTALL_PREFIX=</path/to/third_party/opencv4> ..
      cmake --build .
      make install
      cd ../..
      

      This OpenCV may use Eigen instead of LAPACK as the linear alegbra package.

   3. Download LibTorch.

      wget https://download.pytorch.org/libtorch/cu113/libtorch-cxx11-abi-shared-with-deps-1.12.1%2Bcu113.zip
      unzip libtorch-cxx11-abi-shared-with-deps-1.12.1+cu113.zip
      

   4. Clone and install Torchvision

      git clone https://github.com/pytorch/vision.git && cd vision
      git checkout tags/v0.13.1
      mkdir build && cd build
      cmake -DCMAKE_INSTALL_PREFIX=</path/to/third_party/torchvision> \
            -DCMAKE_PREFIX_PATH=</path/to/third_party/libtorch/share/cmake> \
            -DWITH_CUDA=ON \
            -DUSE_PYTHON=OFF ..
      cmake --build .
      make install
      cd ../..
      

5. Compile the source code

   mkdir build && cd build
   cmake -DCMAKE_BUILD_TYPE=Release ..
   cmake --build .
   

6. Run FairMOT

   ./FairMOT </path/to/video/file>
   

Tracking performance

Results on MOT challenge train set.

Dataset	MOTA	IDF1	IDS	MT	ML	FPS
MOT-16	81.8	80.6	519	303	27	30.1

Evaluation was run on NVIDIA GeForce RTX 3070 8GB.

Demo