Answers are grounded in this paper and cite their sources.
Skeleton-based action recognition supports applications such as human–computer interaction and rehabilitation monitoring, motivating the development of accurate models with compact parameterizations. Although Transformers capture long-range spatiotemporal dependencies, their representations may contain redundant joint and channel responses, uninformative frames, and insufficiently constrained latent statistics. We propose SKD-Net, a lightweight framework that combines semantic and temporal refinement with latent regularization. The Semantic Enhancement Module (SEM) constructs a channel descriptor from a sample-adaptive subset of joints to guide global–local channel recalibration while preserving the complete joint representation. The Key Frame Selection Module (KFSM) combines independently normalized feature-space motion variation and skeletal directional variation to rank and retain frames without an additional learnable temporal scoring network. These designs are complemented by the Discriminative Representation Constraint Module (DRCM), which adopts established first-order Maximum Mean Discrepancy regularization to align global and class-conditional latent means with reference means under classification supervision. Experiments on NTU RGB+D 60, NTU RGB+D 120, and NW-UCLA demonstrate competitive recognition performance with 1.98 million parameters. Repeated-run evaluations show modest average improvements over the reproduced baseline on the two NTU Cross-Subject protocols, while controlled ablations examine the contributions of temporal selection and latent regularization. The results support SKD-Net as a compact recognition framework for researchers and developers interested in skeleton-based action recognition with a limited model parameter budget.
No takes yet. Share an insight, caveat, or question.
Tang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: