Example of missing marker reconstruction

This is an implementation for the paper A Neural Network Approach to Missing Marker Reconstruction in Human Motion Capture. Youtube video of the experimental results is here.

• Python 3
• Tensorflow >= 1.0
• Additional Python libraries:
  • numpy
  • matplotlib (if you want to visualize the results)
  • btk (if you want to preprocess test sequences by yourself)

In my experiments I have been using CMU Mocap dataset. There are 2 options on how to get it:

1. Download already preprocessed dataset:

   Take the test sequences, I used in the paper here

2. Preprocess dataset by yourself

   (** Note: the results depends a lot on amount and type of the data you use **)

   • Install btk library
   • Download CMU Mocap dataset in BVH format
   • Create folders "train" and "eval" (if flag "evaluate" is true) or "dev" (if flag "evaluate" is false)
   • Move the files you want to use for training into the "train" folder (they should be in the folder themself)
   • Move the files you want to use for testig into the "eval" or "dev" folder (they should be in the folder themself)
   • Set the address to this data in the code/ae/utils/flag.py as data_dir
   • Preprocess it by running the script code/ae/utils/data.py

   Afterwards you need to put all test sequences you want to test on into the folder "test_seq", which should be in the same directory as the main folder with the data. Then you preprocess those sequence by function "write_test_seq_in_binary" from the file ae/utils/data.py, which will write the test sequences in the binary format for the faster and easier access.

So final configuration should look like this:

''' .../adress_to_the_data/...

/folder_with_the_data

• eval.binary
• maximums.binary
• mean.binary
• train.binary
• variance.binary

/test_seq

• basketball.binary
• boxing.binary
• salto.binary ... '''

To run the default example execute the following command.

$ python code/ae/train.py

You can play around with the run options, including the neural net size and shape, input corruption, learning rates, etc. in the file flags.py. Otherwise - you can find the Best Flags in the folder BestFlags

If you encounter any problems/bugs/issues please create an issue on github or email me at [email protected] for any bug reports/questions/suggestions. I prefer questions and bug reports on github as that provides visibility to others who might be encountering same issues or who have the same questions.

Here is the citation in bib format:

@article{kucherenko2018neural,
  title={A Neural Network Approach to Missing Marker Reconstruction in Human Motion Capture},
  author={Kucherenko, Taras and Beskow, Jonas and Kjellstr{\"o}m, Hedvig},
  journal={https://arxiv.org/abs/1803.02665v4},
  year={2018}
}