Answers are grounded in this paper and cite their sources.
This paper proposes a Multi-View Spatio-Temporal Transformer (MVST) for skill-aware tennis stroke recognition from paired skeleton sequences.Estimating a player's skill level from movement patterns, beyond recognizing the performed action itself, remains a challenging problem in computer vision.A shared transformer encoder models spatial landmark relationships and temporal motion dynamics, while back-view and side-view representations are fused to reduce viewpoint ambiguity.A Cross-View Alignment (CVA) objective is introduced as an auxiliary contrastive loss that aligns view-level embeddings of the same stroke instance during training, with no inference overhead.On a self-collected dataset of eight subjects and 96 paired stroke instances, evaluated under a reproducible five-seed, four-fold subject-independent protocol, MVST with CVA achieves a final-epoch accuracy of 0.83, showing higher mean accuracy than ST-GCN, SkateFormer and SA-TDGFormer while achieving comparable performance to CTR-GCN with 2.4 times fewer parameters and 1.4 times fewer FLOPs.A leave-one-subject-out analysis achieves 0.85 skill-classification accuracy on subjects excluded from training, indicating that the learned representation can generalize to unseen subjects within this cohort.
No takes yet. Share an insight, caveat, or question.
Asror et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: