This is a pytorch implementation of the paper Periodicity Counting in Videos with Unsupervised Learning of Cyclic Embeddings (preprint version available here). This implementation is oriented for regular videos. To adapt it to other input natures, change the architecture in training_utils/triplet_based_encoder.py so that it fits your input. Here are some examples.

• For 4D MRI, change the 2D convolution layers to 3D ones.
• For complex time series, use MLP layers instead of convolutions.

1. in the file count_periodicities.py, write the path to your video in the variable video_path.
2. define the number of epochs you need for training with the variable epochs_nb.
3. you can change the parameters of the Max Detector algorithm at line 16 when calling the function count_repetitions¹²
4. launch count_periodicities.py

¹: period_range is the maximum variation of duration from a cycle to the next
²: N represents the number of frequencies to evaluate.

If this was useful, please cite the article:

@article{jacquelin:hal-03738161,
  TITLE = {Periodicity Counting in Videos with Unsupervised Learning of Cyclic Embeddings},
  AUTHOR = {Jacquelin, Nicolas and Vuillemot, Romain and Duffner, Stefan},
  URL = {https://hal.archives-ouvertes.fr/hal-03738161},
  JOURNAL = {Pattern Recognition Letters},
  PUBLISHER = {Elsevier},
  YEAR = {2022},
  DOI = {10.1016/j.patrec.2022.07.013},
}