Multimodal Learning inspired by Multitask learning in Video Emotion Recognition
Multimodal Learning inspired by Multitask learning in Video Emotion Recognition
Introduction
This is my on-going project. I use the idea of Multitask learning for Multimodal learning. In our proposed methods, we utilize the idea of multitask learning for multimodal learning. In multitask learning, a prediction of a task can improve the accuracy of the prediction of another task. For multimodal learning, we aim to use a prediction generated from a specific modal to enhance the prediction obtained from another modal or a prediction of the late-fusion production.
Dataset
We conducted the experiments on the reorganized IEMOCAP and CMU-MOSEI provided by Dai, and the CH-SIMS dataset. The preprocessed IEMOCAP and CMU-MOSEI can be downloaded here. The CH-SIMS and its baseline can be found here. For all datasets, we used the MTCNN to extract facial images and resized to 260x260.
Environment
- timm 0.4.5
Results
Test results on IEMOCAP
| Model | Ang_Ac | Ang_F1 | Exc_Ac | Exc_F1 | Fru_Ac | Fru_F1 | Hap_Ac | Hap_F1 | Neu_Ac | Neu_F1 | Sad_Ac | Sad_F1 | Avg_Ac | Avg_F1 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Audio | 82.1 | 46.2 | 55.0 | 30.3 | 45.4 | 42.4 | 12.8 | 17.2 | 48.6 | 41.4 | 77.9 | 48.8 | 53.6 | 37.7 |
| Visual | 80.9 | 53.8 | 84.3 | 56.3 | 68.3 | 53.9 | 90.1 | 43.7 | 75.1 | 55.3 | 86.6 | 55.8 | 80.9 | 53.1 |
| Text | 86.0 | 57.9 | 86.8 | 56.8 | 70.1 | 54.7 | 90.1 | 40.5 | 73.5 | 49.0 | 87.0 | 58.0 | 82.2 | 52.8 |
| LF-LSTM | 71.2 | 49.4 | 79.3 | 57.2 | 68.2 | 51.5 | 67.2 | 37.6 | 66.5 | 47.0 | 78.2 | 54.0 | 71.8 | 49.5 |
| LF-Trans | 81.9 | 50.7 | 85.3 | 57.3 | 60.5 | 49.3 | 85.2 | 37.6 | 72.4 | 49.7 | 87.4 | 57.4 | 78.8 | 50.3 |
| EmoEmbs | 65.9 | 48.9 | 73.5 | 58.3 | 68.5 | 52.0 | 69.6 | 38.3 | 73.6 | 48.7 | 80.8 | 53.0 | 72.0 | 49.8 |
| MulT | 77.9 | 60.7 | 76.9 | 58.0 | 72.4 | 57.0 | 80.0 | 46.8 | 74.9 | 53.7 | 83.5 | 65.4 | 77.6 | 56.9 |
| BIMHA | 77.2 | 57.6 | 78.3 | 56.1 | 73.9 | 54.2 | 83.4 | 43.2 | 76.4 | 50.9 | 83.8 | 63.7 | 78.8 | 54.3 |
| CMHA | 88.6 | 61.1 | 87.9 | 60.5 | 75.1 | 56.3 | 89.0 | 45.8 | 76.5 | 51.2 | 88.3 | 61.6 | 84.3 | 56.1 |
| MESE | 88.2 | 62.8 | 88.3 | 61.2 | 74.9 | 58.4 | 89.5 | 47.3 | 77.0 | 52.0 | 88.6 | 62.2 | 84.4 | 57.4 |
| FE2E | 88.7 | 63.9 | 89.1 | 61.9 | 71.2 | 57.8 | 90.0 | 44.8 | 79.1 | 58.4 | 89.1 | 65.7 | 85.7 | 57.1 |
| Le et al | 90.1 | 66.8 | 88.5 | 66.8 | 77.7 | 57.0 | 90.5 | 48.5 | 78.1 | 56.6 | 90.7 | 69.6 | 85.9 | 60.9 |
| Ours | 89.5 | 67.8 | 90.8 | 70.7 | 78.9 | 59.9 | 89.9 | 55.5 | 79.1 | 60.6 | 91.4 | 72.9 | 86.6 | 64.6 |
Test results on MOSEI
| Model | Ang_Wa | Ang_F1 | Dis_Wa | Dis_F1 | Fea_Wa | Fea_F1 | Hap_Wa | Hap_F1 | Sad_Wa | Sad_F1 | Sur_Wa | Sur_F1 | Sur_Wa | Sur_F1 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Audio | 53.9 | 40.5 | 61.0 | 35.7 | 59.0 | 19.6 | 50.0 | 69.3 | 61.2 | 45.8 | 58.4 | 21.7 | 57.2 | 38.8 |
| Visual | 58.9 | 38.2 | 63.2 | 37.6 | 59.1 | 21.8 | 55.7 | 70.3 | 56.2 | 42.8 | 53.0 | 17.9 | 57.7 | 38.1 |
| Text | 65.9 | 48.4 | 74.0 | 56.0 | 62.8 | 27.0 | 62.3 | 72.0 | 60.2 | 45.3 | 60.9 | 26.0 | 64.3 | 45.8 |
| LF-LSTM | 64.5 | 47.1 | 70.5 | 49.8 | 61.7 | 22.2 | 61.3 | 73.2 | 63.4 | 47.2 | 57.1 | 20.6 | 63.1 | 43.3 |
| LF-Trans | 65.3 | 47.7 | 74.4 | 51.9 | 62.1 | 24.0 | 60.6 | 72.9 | 60.1 | 45.5 | 62.1 | 24.2 | 64.1 | 44.4 |
| EmoEmbs | 66.8 | 49.4 | 69.6 | 48.7 | 63.8 | 23.4 | 61.2 | 71.9 | 60.5 | 47.5 | 63.3 | 24.0 | 64.2 | 44.2 |
| MulT] | 64.9 | 47.5 | 71.6 | 49.3 | 62.9 | 25.3 | 67.2 | 75.4 | 64.0 | 48.3 | 61.4 | 25.6 | 65.4 | 45.2 |
| BIMHA | 65.3 | 47.4 | 70.5 | 48.9 | 61.8 | 24.7 | 65.8 | 72.1 | 62.6 | 47.9 | 62.5 | 24.9 | 64.8 | 44.3 |
| CMHA | 65.9 | 49.1 | 73.6 | 53.2 | 63.4 | 27.3 | 65.2 | 72.1 | 64.2 | 46.7 | 64.5 | 26.6 | 66.1 | 45.8 |
| MESE | 66.8 | 49.3 | 75.6 | 56.4 | 65.8 | 28.9 | 64.1 | 72.3 | 63.0 | 46.6 | 65.7 | 27.2 | 66.8 | 46.8 |
| FE2E | 66.9 | 49.5 | 75.4 | 57.2 | 63.8 | 27.1 | 61.9 | 72.3 | 65.6 | 49.3 | 61.5 | 26.9 | 65.8 | 47.0 |
| Le et al | 67.5 | 50.2 | 76.3 | 57.0 | 69.0 | 29.0 | 63.0 | 72.6 | 65.5 | 49.2 | 65.7 | 27.6 | 67.8 | 47.6 |
| Ours | 71.4 | 52.6 | 81.4 | 57.4 | 80.5 | 30.2 | 68.5 | 75.4 | 63.9 | 51.6 | 80.3 | 30.3 | 74.3 | 49.6 |
Test results on CH-SIMS
| Model | Annotation | Acc2 | F1 | MAE | Corr |
|---|---|---|---|---|---|
| EF-LSTM [30] | M | 69.37 | 81.91 | 59.34 | -4.39 |
| MFN [31] | M | 77.86 | 78.22 | 45.19 | 55.18 |
| MulT [3] | M | 77.94 | 79.10 | 48.45 | 55.94 |
| LF-DNN [28] | M | 79.87 | 80.20 | 42.01 | 61.23 |
| MLF-DNN [27] | M, A, T, V | 82.28 | 82.52 | 40.64 | 67.47 |
| LMF [29] | M | 79.34 | 79.96 | 43.99 | 60.00 |
| MLMF [27] | M, A, T, V | 82.32 | 82.66 | 42.03 | 63.13 |
| TFN [28] | M | 80.66 | 81.62 | 42.52 | 61.18 |
| MTFN [27] | M, A, T, V | 82.45 | 82.56 | 40.66 | 66.98 |
| Self-MM [26] | M, A, T, V | 80.74 | 80.78 | 41.90 | 61.60 |
| Human-MM [26] | M, A, T, V | 81.32 | 81.73 | 40.80 | 64.70 |
| Ours | M, A, T, V | 83.37 | 83.15 | 37.61 | 68.04 |