Table 1. Summary of various models for emotion and intensity recognition.
| Method | Model | DB & performance | Limitation |
|---|---|---|---|
| Verma et al. (2005) | Distance based | Primary source: NA | Only few emotions are considered, method not generalise, emotion intensity before emotion recognition, computationally expensive. |
| Lee & Xu (2003) | Optical flow tracking algorithm (Distance) | Real-time data | Need for each subject to be trained differently, not generalise, predicting intensity before emotion |
| Kim & Pavlovic (2010) | HCORF (Prob) | CMU | Intrinsic topology of FER data is linearly model. |
| Quan, Qian & Ren (2014) | K-Means (Cluster) | CK+ | Predict intensity before emotion, intensity estimation based on graphical difference is not logical |
| Chang, Chen & Hung (2013) | Scatering transform + SVM (Cluster) | CK+ | Emotion recognition task is omitted. |
| Zhao et al. (2016) | SVOR (Regression) | Pain | Correlations between emotion classes are not modelled. |
| Rudovic, Pavlovic & Pantic (2012) | LSM-CORF (Prob) | BU-4DFE, CK+ | Latent states are not considered in the modeling of sequences across and within the classes |
| Walecki et al. (2015) | VSL-CRF (Prob) | CK+ AFEW | Result of emotion intensity is not accounted for. |
| Kamarol et al. (2017) | weighted vote | CK+ | Emotion and emotion intensity not concurrently predicted. |
| Proposed model | ML-CNN (Multi-Label) | BU-3DFE | Assume temporal information among sequence data as ordinal metrics. |
Note:
NA: Not Applicable, MAE: Mean Absolute Error, PCC: Pearson Correlation Coefficient, ICC: Intraclass Correlation, MAL: MeanAbsolute Loss, HL: Hamming Loss, RL: Ranking Loss; AP: Average Precision, CE: Coverage Error.