Table 7.
Performance comparison of the recent methods with the proposed bi-modular sequential neural network.
| Performance Comparison Experiments | ||||||
|---|---|---|---|---|---|---|
| Method | Class AP (angry) | Class AP (happy) | Class AP (sad) | Class AP (neutral) | Macro-mAP | Micro-mAP |
| STEP (2020) [36] | 0.22 | 0.52 | 0.30 | 0.12 | 0.29 | 0.27 |
| ADF (2019) [38] | 0.22 | 0.59 | 0.30 | 0.12 | 0.31 | 0.27 |
| STGCN (2018) [35] | 0.06 | 0.97 | 0.20 | 0.01 | 0.34 | 0.41 |
| HAPAM (2020) [37] | 0.97 | 0.66 | 0.40 | 0.18 | 0.60 | 0.88 |
|
Proposed LSTM
and MLP (RGS) |
0.98 | 0.74 | 0.58 | 0.33 | 0.66 | 0.92 |
|
Proposed LSTM
and MLP (RGS + JRA + JRD) |
0.99 | 0.90 | 0.84 | 0.46 | 0.80 | 0.96 |
|
Proposed LSTM and MLP with
batch normalization (RGS + JRA + JRD) |
0.99 | 0.91 | 0.90 | 0.65 | 0.86 | 0.97 |
The proposed bi-modular networks outperform the previous state-of-the-art methods in mean average precision scores.