Facial Expression Recognition (FER) traditionally relies on raw RGB images, introducing privacy risks, sensitivity to illumination variations, and susceptibility to identity-confounding visual shortcuts. As an alternative, this work proposes a lightweight, privacy-aware pipeline using exclusively geometric representations-facial landmarks (478 points) and blendshapes (52 coefficients, ARKit standard)-eliminating persistent raw video storage. Six neural architectures trained from scratch on fused CREMA-D and RAVDESS datasets were benchmarked against an ImageNet-pretrained ResNet50 baseline under a 5-fold subject-independent cross-validation protocol. Geometric models consistently outperformed the pixel baseline on unseen subjects with 12-124 times fewer parameters, while demonstrating superior robustness on subtle emotional expressions such as sadness. Ablation studies reveal that blendshapes capture the primary discriminative signal, while landmarks provide essential complementary spatial topology. These results demonstrate that compact geometric representations preserve affective semantics, eliminate downstream transfer learning dependency on visual backbones, and enable efficient, privacy-aware emotion recognition.
Figures & tables
Fig. 1: Overview of the proposed emotion classification pipeline: facemesh landmarks ( 478 points) and blendshape coefficients ( 52 weights) are extracted via MediaPipe Face Landmarker while raw RGB frames are discarded and visually abstracted into geometric primitives to prevent biometric exposure.
Architecture
Parameters
Accuracy (%)
Macro F1-Score (%)
ResNet-1D
1,001,158
58.06±6.59
57.65±6.38
Temporal GNN
208,870
57.15±4.06
56.50±4.10
ST-GCN
834,898
56.64±5.45
56.35±5.22
Temporal TCN
2,155,142
56.64±5.69
55.88±6.10
LSTM
1,500,294
55.86±5.88
55.28±5.92
Temporal MLP
973,958
55.05±6.39
54.68±6.24
TABLE I: Subject-independent 5-fold cross-validation test performance and trainable parameters per architecture. Bold values indicate the best results.
Fig. 2: Comparison of per-class accuracy for each architecture.
Variant
Accuracy (%)
Macro F1-Score (%)
Complete model
57.15±4.06
56.50±4.10
Blendshapes only
53.64±5.53
52.64±5.86
Landmarks only
48.56±1.98
47.44±1.40
TABLE II: GNN performance by input modality. Bold values indicate the best results.
Department of Informatics and Telematics, Harokopio University of Athens, Athens, Greece · Research & Innovation Development Department, Netcompany-Intrasoft S.A., Luxembourg, Luxembourg · Department of Informatics, Faculty of Science, Democritus University of Thrace, Kavala, Greece +3