Denoising and completion filters for human motion software

This repository includes implementation codes of filtering methods for denoising and completing data generated by software platforms for human motion analysis, allowing readers to easily reproduce all the algorithms in different experimental settings.

Usage

To run a simple filter from the main directory:

python3 filters/<filter_name>.py <filter_name> <input_file_path> <output_directory_path> <delta_t> <model>

To check the error of the filtered file, run:

python3 utils/checker.py --ref <ground_truth_directory_path> --source <input_directory_path> -out <stats_file_path>

For example, to run and evaluate the error the Butterworth Filter (BF):

python3 filters/BF.py BF data/input/S9_Discussion.csv data/output/ 20
python3 utils/checker.py --ref data/gt --source data/output/BF --out data/stats/BF.csv

This will create a new directory "data/output/BF" and store the refined results.

To have a reference of the input error, run:

python3 utils/checker.py --ref data/gt --source data/input --out data/stats/input.csv

Data format

For the data format, please follow the example file in the "data" folder.

Training

Some filters may require a training phase. To preprare the train and validation datasets, follow the code in "utils/prepare_data.py". For the training phase, follow the Jupyther Notebook "utils/train.ipynb".

Accuracy Results on Human3.6M

Category	Method	MPJPE Low error jitter	Accel Low error jitter	MPJPE High error jitter	Accel High error jitter	MPJPE Missing keypoints	Accel Missing keypoints	MPJPE Missing keypoints with error	Accel Missing keypoints with error
	Baseline	96.1	230.0	146.3	381.9	-	-	-	-
General Purpose	Simple Moving Average	74.6	11.5	85.0	19.7	-	-	-	-
General Purpose	Weighted Moving Average	76.8	4.4	88.7	7.0	-	-	-	-
General Purpose	Exponential Moving Average	55.4	85.8	89.6	140.8	-	-	-	-
General Purpose	HDE	63.0	92.3	102.6	151.4	-	-	-	-
General Purpose	Butterworth	24.8	3.0	41.5	5.0	-	-	-	-
General Purpose	Savitzky–Golay	63.3	116.7	102.1	194.5	-	-	-	-
General Purpose	Interpolation	-	-	-	-	0.3	1.1	244.8	634.0
General Purpose	Least Squares Gaussian	62.7	17.2	78.4	19.4	-	-	-	-
State Observer	Kalman Filter (0th)	46.4	29.4	66.0	48.3	30.1	2.8	48.7	29.6
State Observer	Kalman Filter (1st)	42.3	33.0	65.9	54.2	21.4	2.4	44.2	33.2
State Observer	Kalman Filter (2nd)	43.1	35.7	68.8	58.7	19.0	2.4	45.0	36.0
Dimensionality Reduction	Truncated Singular Value Decomposition	64.0	143.7	98.3	226.6	-	-	-	-
Dimensionality Reduction	Low-Rank Matrix Completion	-	-	-	-	4.9	16.7	92.5	219.7
Dimensionality Reduction	Noisy Low-Rank Matrix Completion	-	-	-	-	32.3	109.4	111.1	267.8
Dimensionality Reduction	Robust Principal Component Analysis	79.8	182.0	90.2	207.7	-	-	-	-
Deep Neural Network	Undercomplete Autoencoder	83.6	197.7	128.0	331.0	-	-	-	-
Deep Neural Network	Undercomplete Denoising Autoencoder	59.5	89.5	92.0	75.6	-	-	-	-
Deep Neural Network	Recurrent Neural Network Autoencoder	89.5	34.8	136.2	40.8	-	-	-	-
Hybrids	KF+DE	76.9	102.2	138.0	174.3	-	-	-	-

Citation

@article{martini2025denoising,
  title={Denoising and completion filters for human motion software: A survey with code},
  author={Martini, Enrico and Calanca, Andrea and Bombieri, Nicola},
  journal={Computer Science Review},
  volume={58},
  pages={100780},
  year={2025},
  publisher={Elsevier}
}