Spatio-temporal motion adaptive graph convolutional network for facial micro-expression recognition

Micro-expressions (MEs) are brief, involuntary facial movements that reveal genuine emotions and are valuable in applications such as lie detection and criminal investigation. Recent advances in micro-expression recognition (MER) have leveraged Graph Neural Networks to model facial sequences as spatio-temporal graphs, often enhanced with attention mechanisms to emphasize informative landmarks and critical frames. However, existing attention designs typically derive importance weights via learned transformations of intermediate features, which are not explicitly conditioned on sequence-specific motion cues. To address this issue, we propose the Spatio-Temporal Motion Adaptive Graph Convolutional Network (ST-MAGCN), which computes sample-specific landmark and frame weights from temporal feature differences, enabling motion-aware adaptation during inference. The core intuition is that facial regions and frames with greater motion variation carry more discriminative and complementary information. Moreover, we introduce complete graph convolution to capture long-range dependencies between facial landmarks, further enhancing the detection of fine-grained facial changes. Experiments on CASME II, SAMM, and SMIC demonstrate that our approach consistently outperforms existing methods and achieves notable improvements in MER.