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.

对于人物重新识别的挑战，本研究以在图像库中识别特定个体为主要目标，改进了现有的重新识别方法，并提出了 AaP-ReID，一种将通道注意力融入 ResNet 架构的人物重新识别更有效的方法。经过在三个基准数据集上的评估，相较于现有方法，在 Market-1501 上达到 95.6% 的准确率，DukeMTMC-reID 上达到 90.6% 的准确率，并在 CUHK03 上达到 82.4% 的准确率。