Scholarlink

<p>Group Activity Recognition (GAR) aims to infer the group activity by modeling the interactions among multiple actors within a scene, and it has broad real-world applications. Under weak supervision, existing methods are limited by rigid feature extraction that fails to adaptively capture salient visual cues and by linear attention mechanisms that discard essential negative interactions due to non-negativity constraints. To address these issues, we propose a Dynamic Polarity Learning Network (DPLN) with three core components: 1) A Multi-dimensional Attention Convolution (MAC) module that enables adaptive feature extraction by learning complementary attention weights in parallel across all four dimensions of the convolutional kernels; 2) a Polar Attention Encoder (PAE) that explicitly models positive and negative query–key interactions to recover overlooked negative information and generate more discriminative token representations; 3) a Convolutional Attention Aggregation (CAA) module that combines 1D convolution with enhanced linear attention and RoPE to efficiently aggregate spatiotemporal features with linear complexity. Under weak supervision, DPLN achieves state-of-the-art performance on the Volleyball and NBA datasets, outperforming existing weakly supervised methods and even surpassing some strongly supervised approaches.</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

Dynamic polarity learning network for weakly supervised group activity recognition

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Dongli Wang,
  • Tao Zhu,
  • Yan Zhou,
  • Xiaolin Zhu,
  • Zixin Zhang,
  • Richard Irampaye

摘要

Group Activity Recognition (GAR) aims to infer the group activity by modeling the interactions among multiple actors within a scene, and it has broad real-world applications. Under weak supervision, existing methods are limited by rigid feature extraction that fails to adaptively capture salient visual cues and by linear attention mechanisms that discard essential negative interactions due to non-negativity constraints. To address these issues, we propose a Dynamic Polarity Learning Network (DPLN) with three core components: 1) A Multi-dimensional Attention Convolution (MAC) module that enables adaptive feature extraction by learning complementary attention weights in parallel across all four dimensions of the convolutional kernels; 2) a Polar Attention Encoder (PAE) that explicitly models positive and negative query–key interactions to recover overlooked negative information and generate more discriminative token representations; 3) a Convolutional Attention Aggregation (CAA) module that combines 1D convolution with enhanced linear attention and RoPE to efficiently aggregate spatiotemporal features with linear complexity. Under weak supervision, DPLN achieves state-of-the-art performance on the Volleyball and NBA datasets, outperforming existing weakly supervised methods and even surpassing some strongly supervised approaches.