Dynamically Gated Similarities

This is the code for the Thesis "Multi-Person Pose Tracking using Dynamically Gated Similarities", available here (or see: ./thesis.pdf).

Documentation

You can find the extended Documentation over at https://bmmtstb.github.io/dynamically-gated-similarities/.

Pipeline

You can find a visual Pipeline on LucidChart or downloadable as PDF (or see: ./docs/figures/Pipeline-DGS-Overview.pdf). The visual pipeline of the training module is also available as PDF (or see: ./docs/figures/Pipeline-DGS-Training.pdf).

Examples

pt21__val__012968_mpii_test__prediction

dance__test__dancetrack0017__prediction

Citing

To cite this thesis, you can use the following BibTeX entry:

@mastersthesis{tuprints29468,
           title = {Multi-Person Pose Tracking Using Dynamically Gated Similarities},
          author = {Martin Steinborn},
          school = {Technische Universit{\"a}t Darmstadt},
        language = {en},
         address = {Darmstadt},
            year = {2025},
           pages = {VII, 56 Seiten},
           month = {M{\"a}rz},
             url = {http://tuprints.ulb.tu-darmstadt.de/29468/},
             doi = {https://doi.org/10.26083/tuprints-00029468},
        keywords = {tracking, pose-tracking, mppt}
}

To cite the code, you can use the following BibTeX entry:

@software{brizar_2025_14910547,
  author       = {Brizar},
  title        = {bmmtstb/dynamically-gated-similarities},
  month        = feb,
  year         = 2025,
  publisher    = {Zenodo},
  version      = {v0.3.0},
  doi          = {10.5281/zenodo.14910547},
  url          = {https://doi.org/10.5281/zenodo.14910547},
}

This code is based on works of other people as well. Please refer to the respective repositories for citing.