Keypoint-based Multi-Object Tracking (keypoint-mot)

This repository is intended for the Multi-Object Tracking as part of the internship project.

It provides an implementation of CenterTrack in tensorflow.

Installation

#clone the repo
git clone https://github.com/googleinterns/keypoint-mot.git
cd keypoint-mot

#optional, create a new virtual environment and activate it
#install the dependencies
pip install -r requirements.t

Next, if this PR was approved, simply install tensorflow addons.

pip install tensorflow-addons #install tensorflow addons

Otherwise, tensorflow addons must be built from the repo in the PR.

git clone https://github.com/Licht-T/addons.git #clone the repo
cd addons
git checkout add-deformable-conv

export TF_NEED_CUDA="1"

# Set these if the below defaults are different on your system
export TF_CUDA_VERSION="10.1"
export TF_CUDNN_VERSION="7"
export CUDA_TOOLKIT_PATH="/usr/local/cuda"
export CUDNN_INSTALL_PATH="/usr/lib/x86_64-linux-gnu"

# This script links project with TensorFlow dependency
python3 ./configure.py

#add the --cxxopt option only if you are using gcc 5 or higher
bazel build --enable_runfiles build_pip_pkg --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"
bazel-bin/build_pip_pkg artifacts

pip install artifacts/tensorflow_addons-*.whl

Usage

To instantiate CenterTrack (DLASeg instance), below is a usual configuration:

heads = {'hm': 10, 'reg': 2, 'wh': 2, 'tracking': 2, 'dep': 1,
         'rot': 8, 'dim': 3, 'amodel_offset': 2}
head_conv = {'hm':  [256], 'reg': [256], 'wh': [256],
             'tracking': [256], 'dep': [256], 'rot': [256],
             'dim': [256], 'amodel_offset': [256]}
opt = DLASegOptions(...)
model = DLASeg(num_layers=34, heads=heads, head_convs=head_conv, opt=opt)

For data loading, there are two options:

• Dataset.from_generator(dataset_instance.get_input_generator(args), output_types=dataset_instance.return_dtypes)
• Dataset.range(dataset_len).map(map_func=dataset_instance.get_input_py_func, num_parallel_calls=tf.data.experimental.AUTOTUNE)

Acknowledgement

The code is based on the official CenterTrack pytorch implementation, released under MIT License. Please see the NOTICE for details.

This is not an officially supported Google product.