Table 7 Recognition accuracy comparison for the MSR-Action3D dataset
From: Flexible human action recognition in depth video sequences using masked joint trajectories
Method | Accuracy | Type |
|---|---|---|
Proposed method | 84.09 % | Skeleton |
Proposed method (no noise masking) | 79.31 % | Skeleton |
Rate-invariant analysis (NN) [37] | 63 % | Skeleton |
Dynamic temporal warping [34] | 54 % | Skeleton |
MMTW [35] | 92.57 % | Skeleton |
Joint movement similarities [36] | 91.2 % | Skeleton |
HOPC [27] | 90.9 % | Depth |
Rate-invariant analysis (SVM) [37] | 89 % | Skeleton |
HON4D [26] | 88.36 % | Depth |
Mining actionlet ensemble [9] | 88.2 % | Skeleton |
Histograms of 3D joints [8] | 78.97 % | Skeleton |
Action graph on bag-of-3D points [24] | 74.7 % | Depth |
Hidden Markov model [29] | 63 % | Skeleton |
Recurrent neural network [46] | 42.5 % | Skeleton |