mocap-refinement

This repository includes implementation codes or links to the authors’ original 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.

Methods

Classical Methods

• ✅ Simple Moving Average (SMA)
• ✅ Weighted Moving Average (WMA)
• ✅ Exponential Moving Average (EMA)
• ✅ Holt Double Exponential Smoothing Filter (HDE)
• ✅ Butterworth (BF)
• ✅ Least Square Gaussian (LSG)
• ✅ Savitzky–Golay (SG)
• 🔄 Interpolation (INT)

State Observers

• ✅ Kalman Filter:
  • ✅ with random walk motion model (KF0)
  • ✅ 1th-order (KF1)
  • ✅ 2th-order (KF2)
• 🔄 Extended Kalman Filter (EKF)
• 🔄 Unscented Kalman Filter (UKF)
• 🔄 Tobit Kalman Filter (TKF)
• 🔗 Bolero-Dynammo

Dimensionality Reduction

• ✅ Truncated Singular Value Decomposition (TSVD)
• ✅ Principal Component Analysis (PCA)
• 🔄 Low-Rank Matrix Completion (LRMC)
• 🔄 Noisy Low-Rank Matrix Completion (NLRMC)
• 🔄 Robust Principal Component Analysis (RPCA)
• 🔄 Non-negative Matrix Factorization (NMF)
• 🔄 Dictionary Learning (DL)

Neural Networks

• 🔗 Autoencoder (Smoothnet - SN)
• 🔗 FLK

Hybrid Approaches

• ✅ Kalman Filter + Differential Evolutionary (Das2017 - KF+DE)

Accuracy Results on Human3.6M

Method	denoising		completion		recovery
	MPJPE (mm)	Accel (mm/s2)	MPJPE (mm)	Accel (mm/s2)	MPJPE (mm)	Accel (mm/s2)
baseline	159.59	390.78	-	-	-	-
SMA	58.73	16.74	-	-	-	-
WMA	44.77	12.68	-	-	-	-
EMA	92.83	145.66	-	-	-	-
HDE	105.88	156.72	-	-	-	-
Butterworth	56.78	15.46	-	-	-	-
Least-Squared	55.44	16.17	-	-	-	-
Savitzky-Golay	74.08	49.99	-	-	-	-
Interpolation	-	-	0.02	0.09	162.18	350.38
KF0	64.81	49.96	28.97	2.14	67.01	49.45
KF1	64.28	53.73	24.24	2.02	66.42	53.26
KF2	65.8	56.52	22.67	2.03	68.05	56.14
T-SVD	111.42	106.33	-	-	-	-
PCA	-	-
KF+DE	131.9	185.76	-	-	-	-