This repository is the official code implementation of the GCNNMatch: Graph Convolutional Neural Networks for Multi-Object Tracking via Sinkhorn Normalization on IEEE and on Arxiv. Link to access a new traffic vehicle monitoring dataset named "VA Beach Traffic Dataset" will be provided here.
If you find this paper or code useful, please cite using the following on IEEE:
@inproceedings{papakis2021graph,
title={A Graph Convolutional Neural Network Based Approach for Traffic Monitoring Using Augmented Detections with Optical Flow},
author={Papakis, Ioannis and Sarkar, Abhijit and Karpatne, Anuj},
booktitle={2021 IEEE International Intelligent Transportation Systems Conference (ITSC)},
pages={2980--2986},
year={2021},
organization={IEEE}
}
or on Arxiv:
@article{papakis2020gcnnmatch,
title={GCNNMatch: Graph Convolutional Neural Networks for Multi-Object Tracking via Sinkhorn Normalization},
author={Papakis, Ioannis and Sarkar, Abhijit and Karpatne, Anuj},
journal={arXiv preprint arXiv:2010.00067},
year={2020}
}
-
Install singularity following instructions from its website.
-
Git clone this repo folder and cd to it.
-
"sudo singularity build geometric.sif singularity". Follow instructions from pytorch-geometric to change settings if needed for your system.
-
Download MOT17 Dataset from MOT website and place it in a folder /MOT_dataset.
-
"mkdir overlay". It will allow you to install additional packages if needed in the future.
-
"sudo singularity run --nv -B /MOT_dataset/:/data --overlay overlay/ geometric.sif"
-
"./create_folders.sh"
-
Command: ./train.sh
-
Result: Training will start and save the trained models in /models. Settings can be changed in tracking.py.
-
Specify which trained model to use in tracking.py. A trained model can be found here.
-
Command: ./test.sh
-
Result: Testing will start and produce txt files and videos saved in /output. Settings can be changed in tracking.py
For Benchmark evaluation the pre-processed with Tracktor detection files from this repo were used.