This paper addresses the problem of pedestrian retrieval, also known as person re-identification (re-ID), by focusing on learning discriminative part-informed features. The authors propose two main contributions: (1) a network named Part-based Convolutional Baseline (PCB) that learns part-level features through uniform partitioning, and (2) a refined part pooling (RPP) method to enhance within-part consistency. PCB takes an image input and outputs a convolutional descriptor consisting of several part-level features. It achieves competitive results with state-of-the-art methods, serving as a strong convolutional baseline for person retrieval. RPP addresses the issue of outliers in each part, which are more similar to other parts, by re-assigning these outliers to the closest parts, thereby improving within-part consistency. Experimental results on the Market-1501 dataset demonstrate that PCB+RPP achieves (77.4±4.2)% mAP and (92.3±1.5)% rank-1 accuracy, surpassing existing methods by a significant margin. The paper also discusses related work, including hand-crafted and deep learning approaches for part features, and compares the proposed methods with state-of-the-art techniques. The authors conclude that their approach, combining PCB and RPP, sets a new standard for person retrieval tasks.This paper addresses the problem of pedestrian retrieval, also known as person re-identification (re-ID), by focusing on learning discriminative part-informed features. The authors propose two main contributions: (1) a network named Part-based Convolutional Baseline (PCB) that learns part-level features through uniform partitioning, and (2) a refined part pooling (RPP) method to enhance within-part consistency. PCB takes an image input and outputs a convolutional descriptor consisting of several part-level features. It achieves competitive results with state-of-the-art methods, serving as a strong convolutional baseline for person retrieval. RPP addresses the issue of outliers in each part, which are more similar to other parts, by re-assigning these outliers to the closest parts, thereby improving within-part consistency. Experimental results on the Market-1501 dataset demonstrate that PCB+RPP achieves (77.4±4.2)% mAP and (92.3±1.5)% rank-1 accuracy, surpassing existing methods by a significant margin. The paper also discusses related work, including hand-crafted and deep learning approaches for part features, and compares the proposed methods with state-of-the-art techniques. The authors conclude that their approach, combining PCB and RPP, sets a new standard for person retrieval tasks.