This repository contains Python code for the algorithm described in the paper A-MAL: Automatic Motion Assessment Learning from Properly Performed Motions in 3D Skeleton Videos, by Tal Hakim and Ilan Shimshoni.

Please cite the paper if you have found A-MAL helpful for your research, as:

@article{hakim2019amal,
  title={A-MAL: Automatic Motion Assessment Learning from Properly Performed Motions in 3D Skeleton Videos},
  author={Hakim, Tal and Shimshoni, Ilan},
  journal={arXiv preprint arXiv:1907.10004},
  year={2019}
}

Note: We are checking the legality options for making the dataset demonstrated in the paper public. For now, the code is available for usage on user data.

• Python 3
• Numpy and Scipy
• SKVP-Python
• simpledtw (optional)

The solution uses SKVP formatted skeleton 3D video files. The SKVP format is detailed here.

The training algorithm requires a directory of SKVP formatted videos of properly-performed motions of the same type. Given such a directory, the training command is:

python run_amal.py train <input directory of videos> <output learned model file> [warping method] [ablations] [reference video output path]

Example: python run_amal.py train /path/to/my_training_videos /path/to/trained_model.txt

• The warping method default value is "our", as described in the paper. It can be overridden by "none" and "dtw". Using the "dtw" option will require installation of the simpledtw module.

• The ablations default value is "none". It can be overridden using one or more ablation types, separated by commas. The possible ablations are: none, active, segmentation, time, diminish. For example, a possible ablations argumemt value can be active,time.

• The reference video output path is only useful when warping with "dtw" method, as the assessment flow will need a reference video to be aligned to (while our default warping method does not need one, as it relies on pre-computed points of interest).

The assessment algorithm requires a pretrained model file and an SKVP formatted video of a motion of the same type as used for training the model. Given these two inputs, the assessment command is:

python run_amal.py test <input model file> <input video file> [warping method] [ablations] [input reference video file]

Example: python run_amal.py test /path/to/trained_model.txt /path/to/a_video_i_want_to_assess.skvp

FEEDBACK: The angle between the right shoulder and the right wrist, through the right elbow, is too acute, from 0.00s to 1.87s
FEEDBACK: The right hand tip is too left, from 0.60s to 1.80s
FEEDBACK: The right hand is too left, from 0.67s to 1.73s
FEEDBACK: The right wrist is too left, from 0.67s to 1.73s
Scores: Active: 0.784891, NonActive: 0.757188, Time: 1.000000
Score: 0.838114

According to the above output, the final score is 83.8%. The score components are listed as well, which refer to the active-joint related, inactive-joint related and time-related parameters. In addition, 4 feedback items are printed.