View-Invariant, Occlusion-Robust Probabilistic Embedding for Human Pose

[AtomScott]

• 1. Overview (basic ideas)

• 2. Novelty

  • Develops a framework for mapping 2D poses to probabilistic embeddings where
  • (1) 2D pose embedding distances correspond to their similarities in absolute 3D pose space
  • (2) a single model embeds different pose visibility patterns, such as those
    from occlusions
  • (3) probabilistic embeddings capture input ambiguity.

• 3. Method (Technical details)

  • 

• 4. Results

  • Embedding performance is measured on action recognition, cross-view pose retrieval and video alignment.
  • ****
  • 

• 5. links to papers, codes, etc.

  • Arxiv: https://arxiv.org/abs/2010.13321

• 6. Thoughts, Comments

  • This paper has showed that it is important to incorporate many tasks in order to learn robust pose embeddings. I don't think there is a better way to learn pose embeddings since the human body has a very complex structure.

• 7. bibtex

  • @misc{liu2020viewinvariant,
    • title={View-Invariant, Occlusion-Robust Probabilistic Embedding for Human Pose},
    • author={Ting Liu and Jennifer J. Sun and Long Zhao and Jiaping Zhao and Liangzhe Yuan and Yuxiao Wang and Liang-Chieh Chen and Florian Schroff and Hartwig Adam},
    • year={2020},
    • eprint={2010.13321},
    • archivePrefix={arXiv},
    • primaryClass={cs.CV}
  • }

• 8. Related Papers