A New Multi-Domain Benchmark for Micro-Action Recognition and Detection

Researchers have introduced MMA-82, a multi-domain benchmark designed to advance the recognition and detection of micro-actions — subtle, short-duration body movements that can signal latent intentions and emotional states, expanding on a prior dataset known as MA-52 [1]. The new benchmark, detailed in a paper submitted to arXiv on June 12, 2026, increases the number of fine-grained micro-action categories from 52 to 82 and contains 77,856 annotated instances drawn from 454 subjects [1][2]. The data spans four distinct domains: laboratory interviews, street interviews, psychiatric patient interviews, and emotion-rich television videos [1][2]. The authors establish two core tasks on MMA-82: Micro-Action Recognition and Multi-label Micro-Action Detection [2]. For recognition, they define both in-domain and cross-domain protocols, including few-shot and zero-shot settings, to test model robustness and generalization [2]. Micro-actions are defined as low-amplitude, whole-body movements that can reveal involuntary reactions and fine-grained affective changes [2]. The previous MA-52 benchmark provided a foundation for this work but was limited in scale, scene diversity, and evaluation protocols [1][2]. The MMA-82 release aims to push micro-action analysis toward more realistic settings [2]. Extensive experiments reported in the paper indicate that current computer vision methods still struggle with realistic micro-action understanding, particularly when facing domain shift, long-tailed category distributions, and complex temporal localization [2]. Beyond benchmarking, the research explores the link between micro-actions and emotion, finding that micro-actions are strongly associated with emotional states and can provide complementary cues to facial micro-expressions, improving emotion recognition performance [2]. The benchmark is publicly available through a project website [1][2]. The work contributes to the broader field of human-centered artificial intelligence, which encompasses capabilities such as perception, reasoning, and decision-making in computational systems [6]. The MMA-82 dataset and associated tasks are designed to serve as a challenging resource for developing and evaluating machine learning models that interpret subtle human behavior [2][5].