Implementation of ICCV21 paper Making Higher Order MOT Scalable: An Efficient Approximate Solver for Lifted Disjoint Paths ( Andrea Hornakova, Timo Kaiser, Paul Swoboda Michal Rolinek, Bodo Rosenhahn, Roberto Henschel )
The introduced approximate solver without tracking framework can be found here!
We use anaconda. Please run the follwing commands to initiate the repository:
git clone https://github.com/TimoK93/ApLift.git
cd ApLift
conda create -n aplift python=3.6
conda activate apliftRun the following commands to install repositories and the approximate solver:
pip install torch==1.6.0
pip install torchvision
pip install numpy
pip install opencv-python scipy attrdict motmetrics tqdm pyyaml matplotlib
pip install git+https://github.com/AndreaHor/LPMP.git@afb5b1bb0564732c6f5fac5aca247a4fca5cf066To run the framework, some data needs to be prepared. Follow the instructions bellow or run the data preparation script.
python prepare_datasets.py Our pretrained classifiers are available here. To use them, unzip it into ./data/.
The datasets used in the paper are provided by www.motchallenge.net. For the evaluation the original data is preprocessed by Tracktor as described in the paper.
Additional to the datasets our tracker needs visual features generated by DG-Net and optical flow between consecutive frames in video sequences.
We wrapped the required features and datasets for the original and preprocessed datasets into ZIP archives. Download is available at:
After successfull download, unzip the files to /data/tmp/.
You can run the training/inference pipeline using scripts main.py and run_full_dataset.py.
To train a pipeline on a set of sequences and evaluate it on a sequence, a configuration file is needed. You can find configuration files in ./config. Please update the configs to your desire. The pipeline is then started by running main.py
python main.py .config/your_slected_config.yamlThe results are stored into a working directory e.g. ./results/some_dir. You can use a preprained model by copying the model checkpoint into the working directory before starting the pipeline!
If you want to run the pipeline for a whole dataset, you can use the script run_full_dataset.py which calls main.py with all sequence combinations for a leave-one-out validation.
python run_full_dataset.py .config/your_slected_config.yamlThe training pipeline does not use any post-processing. To apply post-processing, use one of the following commands:
python3 run_postprocessing.py --working_directory ./result/MyExampleWorkingDirectory --challenge MOT17
python3 run_postprocessing.py --working_directory ./result/MyExampleWorkingDirectory2 --challenge MOT20
After training and inference you can evaluate the results in a working directory using the following command:
python3 evaluate.py --working_directory ./result/MyExampleWorkingDirectory --challenge MOT17The solver performance on crowded scenes (MOT20) is archieved by parallel computation of sub-problems of the global problem instance. For the paper, we solved the problems in a big computation cluster in parallel. While this is not available for everyone, the parallel cluster computing is removed in this repository. Here, the solver solves up to 4 sub-problems in parallel on the computing machine.
To speed up the solver with your cluster, please look into the file ./src/LiftedSolver.py and modify the Processor class.
If you use our work in your research, please cite our publication:
@inproceedings { HorKai2021,
author = {Andrea Hornakova* and Timo Kaiser* and Michal Rolinek and Bodo Rosenhahn and Paul Swoboda and Roberto Henschel and (* equal contribution)},
title = {Making Higher Order MOT Scalable: An Efficient Approximate Solver for Lifted Disjoint Paths},
booktitle = {International Conference on Computer Vision (ICCV)},
year = {2021},
month = {Oct}
}