Person re-identification (ReID) is a well-known problem in the field of computer vision. The primary objective is to identify a specific individual within a gallery of images. However, this task is challenging due to various factors, such as pose variations, illumination changes, obstructions, and the presence ofconfusing backgrounds. Existing ReID methods often fail to capture discriminative features (e.g., head, shoes, backpacks) and instead capture irrelevant features when the target is occluded. Motivated by the success of part-based and attention-based ReID methods, we improve AlignedReID++ and present AaP-ReID, a more effective method for person ReID that incorporates channel-wise attention into a ResNet-based architecture. Our method incorporates the Channel-Wise Attention Bottleneck (CWAbottleneck) block and can learn discriminating features by dynamically adjusting the importance ofeach channel in the feature maps. We evaluated Aap-ReID on three benchmark datasets: Market-1501, DukeMTMC-reID, and CUHK03. When compared with state-of-the-art person ReID methods, we achieve competitive results with rank-1 accuracies of 95.6% on Market-1501, 90.6% on DukeMTMC-reID, and 82.4% on CUHK03.

翻译：行人重识别是计算机视觉领域的一项经典问题，其核心目标是在图像库中识别特定个体。然而，由于姿态变化、光照差异、遮挡以及存在干扰背景等多种因素，该任务极具挑战性。现有行人重识别方法在目标被遮挡时，往往无法捕获判别性特征（如头部、鞋子、背包），反而提取了无关特征。受基于部件和注意力机制的行人重识别方法成功的启发，我们改进了AlignedReID++，提出了一种更有效的行人重识别方法AaP-ReID，该方法将通道注意力机制融入基于ResNet的架构中。我们的方法引入了通道注意力瓶颈（Channel-Wise Attention Bottleneck, CWAbottleneck）模块，通过动态调整特征图中每个通道的重要性来学习判别性特征。我们在三个基准数据集Market-1501、DukeMTMC-reID和CUHK03上评估了AaP-ReID。与当前最先进的行人重识别方法相比，我们取得了具有竞争力的结果，在Market-1501、DukeMTMC-reID和CUHK03数据集上的rank-1准确率分别达到95.6%、90.6%和82.4%。