Abstract
Body-based emotion recognition is crucial for real-time affective systems; however, graph-based skeleton models can be computationally expensive. This paper investigates whether lightweight temporal convolutional networks (TCNs) can serve as an efficient and interpretable alternative for body-based emotion classification.
We evaluate a family of TCN models on the DIEM-A dataset and compare them with a graph-based time-series graph (G-TSG) baseline using metrics such as accuracy, macro-F1, parameter count, and inference latency. Although G-TSG achieves the highest mean performance, TCN-Base remains within $1.58$ accuracy points and $1.25$ macro-F1 points while utilizing $79.18\frac{\text{\%}}{}$ fewer parameters and reducing classifier latency by approximately $12.5\times$.
Additionally, we analyze body-region contributions using region-specific TCN models, zero-based occlusion, and G-TSG gradient saliency. The results indicate that upper-body motion provides the strongest standalone regional cue, the usefulness of body regions varies across emotions, and different interpretability methods capture distinct aspects of model behavior. These findings suggest that lightweight TCNs can effectively support body-based emotion recognition while providing practical insights into how motion cues contribute to classification.
Blogger's Review: The application of lightweight temporal convolutional networks in body-based emotion recognition showcases their potential in efficiency and interpretability, especially in real-time systems. By reducing parameters and latency, TCNs maintain model performance while offering greater practicality and understanding, paving the way for future research in emotion recognition.