高级检索

双路注意力机制行人重识别方法

Person Re-Identification via Dual-Path Attention

  • 摘要: 为解决目前Re-ID方法中对非显著可辨别特征关注不足, 以及提取的行人关键特征表达不充分的问题, 提出一种基于双路注意力机制特征提取网络, 由双路注意力主干网络和增强注意特征融合模块组成. 其中, 双路注意力网络使模型关注到不同显著程度的有效特征区域, 可分别用于挖掘显著和潜在非显著可辨别特征, 强调潜在关键特征的重要性; 增强注意特征融合模块用于完成特征信息互补, 同时采用反事实干预强化习得注意力特征图的质量和有效性,从而得到更具有判别性的最终特征表示. 在 Market1501, DukeMTMC-reID 和 MSMT17 数据集上进行了广泛实验, 结果表明, mAP 值分别达到了 89.3%, 80.0%, 58.4%; Rank-1 值分别达到了 95.7%, 89.8%, 80.7%, 充分证明了该方法的优越性.

     

    Abstract: In order to solve the problem of insufficient attention to non-salient distinguishable features and insufficient expression of extracted key features of pedestrians in current Re-ID methods, a feature extraction network based on a dual-path attention mechanism is proposed (Dual-Path Attention Network), which consists of a dual-path attention backbone network. It is composed of an enhanced attention feature fusion module. Among them, the dual-path attention network allows the model to focus on effective feature areas with different salience levels, which can be used to mine significant and potentially non-salient distinguishable features respectively, emphasizing the importance of potential key features.The enhanced attention feature fusion module is used to complete feature information complementation, and counterfactual intervention is used to enhance the quality and effectiveness of the learned attention feature map, so as to make the final feature representation more discriminative. Extensive experiments were conducted on Market1501, DukeMTMC-reID and MSMT17 data sets, and the results showed that the mAP values reached 89.3%, 80.0%, 58.4% respectively; the Rank-1 values reached 95.7%, 89.8%, 80.7% respectively, which fully proved the superiority of this method.

     

/

返回文章
返回