Unlike obvious actions such as jumping or waving, micro-gestures include slight nods, small hand shifts, or almost imperceptible leg movements – often unconscious signals that reflect true emotional states. Psychological research suggests these behaviors provide more reliable insights than facial expressions or carefully constructed speech. Yet most existing action recognition models, including standard convolutional neural networks (CNNs) and vision transformers, struggle with such fine-grained motion. CNNs have limited temporal sensitivity, while transformers suffer from high computational costs and background noise distractions. Based on these challenges, a dedicated motion-aware architecture for micro-gesture recognition is urgently needed.
A team from the Lappeenranta-Lahti University of Technology (LUT), Finland, and Brno University of Technology, Czech Republic, publishes (DOI: 10.1007/s11633-025-1587-8) their work on April 30, 2026, in the journal Machine Intelligence Research. They introduce Micro-Gesture Mamba-Inspired Linear Attention (MGMILA) – a linear-complexity framework that efficiently captures subtle, short-duration body movements, achieving state-of-the-art results on three benchmark datasets.
MGMILA integrates a Mamba-inspired linear attention (MILA) module, which reduces computational complexity from quadratic to linear, making long video sequences practical. The core innovation lies in three motion-extraction variants: Motion as Layer (MAL), Motion as Content (MAC), and Motion as Gate (MAG). Among these, MAG performs best by using motion features as a gating mechanism – it selectively amplifies gesture-relevant information without disrupting pretrained features. The model also learns to predict human segmentation masks as an auxiliary task, helping it ignore background clutter and focus on body regions. Experiments on three public datasets – iMiGUE (Identity-free Micro-Gesture Understanding and Emotion) , Spontaneous Micro-Gesture (SMG) , and MA-52 – show consistent state-of-the-art performance. On iMiGUE, MGMILA achieves 64.92% Top-1 accuracy, outperforming previous CNN- and transformer-based models. Ablation studies reveal that combining both spatial and temporal motion paths gives the best results, while adding segmentation masks further improves accuracy by nearly 0.7%. Higher input resolution also brings modest gains, confirming that fine spatial details matter for micro-gesture recognition.
“Micro-gestures happen in a blink – sometimes in less than a tenth of a second,” the authors said. “Most AI models simply miss them because they are not designed for such low-intensity, rapid movements. We found that treating motion as a gating mechanism – rather than just an extra layer – helps the model focus naturally on what matters. The human segmentation mask was an unexpected but powerful addition; it pushes the network to look at people, not background noise. This is not just an incremental improvement – it is a different way of thinking about gesture recognition.”
The ability to reliably recognize micro-gestures could transform several fields. In medical diagnostics, subtle body movements may help detect early signs of neurological disorders or psychological distress. In smart vehicles, monitoring a driver's unconscious gestures could improve safety by identifying fatigue or distraction before an accident happens. Sports performance analysis stands to benefit as well – coaches could interpret athletes' genuine emotional responses to competition outcomes, enabling more personalized mental training. Virtual reality systems could also become more intuitive by responding to nearly invisible user commands. Because MGMILA works with standard video and does not require specialized sensors, it is practical for real-world deployment. The team hopes their motion-aware design will inspire a new generation of artificial intelligence (AI) systems that truly understand the human body's quietest language.
###
References
DOI
10.1007/s11633-025-1587-8
Original Source URL
https://doi.org/10.1007/s11633-025-1587-8
Funding information
Open access funding provided by Lappeenranta-Lahti University of Technology LUT, Finland
About Machine Intelligence Research
Machine Intelligence Research (original title: International Journal of Automation and Computing) is published by Springer and sponsored by the Institute of Automation, Chinese Academy of Sciences. The journal publishes high-quality papers on original theoretical and experimental research, targets special issues on emerging topics, and strives to bridge the gap between theoretical research and practical applications.