Scholarlink

<p>Few-Shot Action Recognition (FSAR) enables rapid generalization with minimal annotations, yet existing approaches often struggle to balance computational efficiency with complex temporal modeling. Moreover, spatial misalignments arising from varied viewpoints and execution speeds have been insufficiently explored. In this work, we present Dual-Order Geometric Alignment (DOGA), a framework that adapts the Structured State Space Duality (SSD) theory for robust spatiotemporal alignment. Rather than treating motion implicitly, our Dual-Order Kinematic Prior (DOKP) integrates gated SSD units with second-order acceleration differentials to suppress background noise and explicitly model continuous action evolution. To handle spatial discrepancies, we introduce a Task-Aware Geometric Alignment (TAGA) module that employs interactive spatial contrastive SSDs to guide non-rigid deformation, effectively addressing view drifts. Additionally, we reformulate the classification metric using an Adaptive Bi-directional Soft Matching scheme, which mitigates mean prototype bias via robust between set distance measurement. Systematic experiments on four standard benchmarks demonstrate that DOGA achieves state-of-the-art performance on HMDB51 and UCF101 and competitive results on Kinetics and SSv2, while outperforming other tuple-based methods across evaluated benchmarks in terms of both accuracy and inference efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-order evolution and adaptive geometric alignment via state spaces duality for few-shot action recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月29日

查看全文
  • Shihang Qin,
  • Yang Xu,
  • Feng Yan,
  • Wei Chen,
  • Baowei Zhang,
  • Xiangman Wu,
  • Yiming Gong

摘要

Few-Shot Action Recognition (FSAR) enables rapid generalization with minimal annotations, yet existing approaches often struggle to balance computational efficiency with complex temporal modeling. Moreover, spatial misalignments arising from varied viewpoints and execution speeds have been insufficiently explored. In this work, we present Dual-Order Geometric Alignment (DOGA), a framework that adapts the Structured State Space Duality (SSD) theory for robust spatiotemporal alignment. Rather than treating motion implicitly, our Dual-Order Kinematic Prior (DOKP) integrates gated SSD units with second-order acceleration differentials to suppress background noise and explicitly model continuous action evolution. To handle spatial discrepancies, we introduce a Task-Aware Geometric Alignment (TAGA) module that employs interactive spatial contrastive SSDs to guide non-rigid deformation, effectively addressing view drifts. Additionally, we reformulate the classification metric using an Adaptive Bi-directional Soft Matching scheme, which mitigates mean prototype bias via robust between set distance measurement. Systematic experiments on four standard benchmarks demonstrate that DOGA achieves state-of-the-art performance on HMDB51 and UCF101 and competitive results on Kinetics and SSv2, while outperforming other tuple-based methods across evaluated benchmarks in terms of both accuracy and inference efficiency.