- SVM: 42.41% Accuracy
- Random Forest: 40.08% Accuracy
- SVM/Random Forest: ~10% Accuracy
- Highest Accuracy is achieved using all features (double Landmarks, both FAUs) however probably overfititng (around 300 features)
- SVM RBF reaches 36.58%
- RF reaches 33.33%
- SVM Linear 38.78%, RBF 42.28%
- RF 42.67%
I extracted the landmark distances from the landmark coordinates but this means a feature dimension of +2000 just for the distances, reached 41.25% with SVM (linear), 38.52% with RF
- SVM RBF 43.06%
- SVM Linear 44.36%
- MLP SGD 42.67%
Extracted the PDM parameters using OpenFace, reaching new highest accuracies with only 75 features.
- SVM 45%
- MLP SGD 45.78%
- Standardized 3D landmarks in terms of rotation and positioned, reached 46.17% with SVM Linear in combination with FAU intensity and nonrigid-shape (NO STANDARDSCALER)
| Model | Accuracy (LFW) | EmoRec Accuracy (AffectNet) | Embedding Size |
|---|---|---|---|
| VGG-Face | 98.9% | 34.88% | 4096 |
| Facenet | 99.2% | 32.00% | 128 |
| Facenet512 | 99.6% | 26.38% | 512 |
| OpenFace | 92.9% | 26.12% | 129 |
| DeepFace (FB) | 97.35% | 32.38% | 4096 |
| DeepID | 97.4% | 20.75% | 160 |
| ArcFace | 99.5% | 32.00% | 512 |
| SFace | 99.5% | 32.75% | 128 |
| Seems to be a positive correlation between how a model performs in FaceRec and how well its extracted embeddings perform in EmoRec |