Table 2.
Comparison of our method with state-of-the-art action recognition methods using supervised pose, unsupervised RGB+D, and unsupervised pose on the NTU-60 dataset.
| Method | NTU-60 (CSub) | NTU-60 (CView) | NTU-120 (CSub) | NTU-120 (CSet) |
|---|---|---|---|---|
| Supervised pose-based | ||||
| HOPC [31] | 50.1% | 52.8% | - | - |
| HBRNN [20] | 59.1% | 64.0% | - | - |
| P-LSTM [32] | 62.9% | 70.3% | 25.5% | 26.3% |
| Soft RNN [33] | - | - | 36.3% | 44.9% |
| ST-LSTM [34] | 69.2% | 77.7% | 55.7% | 57.9% |
| VA-RNN-Aug [35] | 79.4% | 87.6% | - | - |
| ST-GCN [26] | 81.5% | 88.3% | - | - |
| IndRNN [36] | 81.8% | 88.0% | - | - |
| HCN [37] | 86.5% | 91.1% | - | - |
| PEM [38] | - | - | 64.6% | 66.9% |
| AS-GCN [11] | 86.8% | 94.2% | - | - |
| ST-GR [39] | 86.9% | 92.3% | - | - |
| DGNN [40] | 87.5% | 94.3% | - | - |
| 2s-AGCN [27] | 88.5% | 95.1% | 82.9% | 84.9% |
| AGC-LSTM [41] | 89.2% | 95.0% | - | - |
| MS-G3D [42] | 91.5% | 96.2% | 86.9% | 88.4% |
| Unsupervised RGBD-based | ||||
| Shuffle and learn [43] | 46.2% | 40.9% | - | - |
| Luo et al. [44] | 61.4% | 53.2% | - | - |
| Li et al. [45] | 68.1% | 63.9% | - | - |
| Unsupervised pose-based | ||||
| LongT GAN [14] | 39.1% | 48.1% | - | - |
| CAE* [30] | - | - | 48.3% | 49.2% |
| P&C FW-AEC [16] | 50.7% | 76.1% | - | - |
| MS2L [28] | 52.6% | - | - | - |
| PDF-G (ours) | 59.7% | 81.0% | 48.2% | 50.9% |
| PDF-G* (ours) | 60.4% | 81.5% | 48.5% | 51.3% |