Abstract
Multimodal contrastive learning has achieved significant performance advantages in self-supervised skeleton-based action recognition. Previous methods are limited by modality imbalance, which reduces alignment accuracy and makes it difficult to combine important spatial-temporal frequency patterns, leading to confusion between modalities and weaker feature representations. To overcome these problems, we explore intra-modality feature-wise self-similarity and inter-modality instance-wise cross-consistency, and discover two inherent correlations that benefit recognition: (i) Global Perspective expresses how action semantics carry a broad and high-level understanding, which supports the use of globally discriminative feature representations. (ii) Focus Adaptation refers to the role of the frequency spectrum in guiding attention toward key joints by emphasizing compact and salient signal patterns. Building upon these insights, we propose a novel language-skeleton contrastive learning framework comprising two key components: (a) Feature Modulation, which constructs a skeleton-language action conceptual domain to minimize the expected information gain between vision and language modalities. (b) Frequency Feature Learning, which introduces a Frequency-domain Spatial-Temporal block (FreST) that focuses on sparse key human joints in the frequency domain with compact signal energy. Extensive experiments demonstrate the effectiveness of our method achieves remarkable action recognition performance on widely used benchmark datasets, including NTU RGB+D 60 and NTU RGB+D 120. Especially on the challenging PKU-MMD dataset, MICA has achieved at least a 4.6% improvement over classical methods such as CrosSCLR and AimCLR, effectively demonstrating its ability to capture internal and contextual attention information.