Pose-guided methods were once dominant in occluded person Re-Identification (ReID) but have received limited attention in recent years. Criticized for their computational overhead and susceptibility to inaccurate pose estimation, particularly in crowded scenes accompanied by low image quality. These methods have also struggled to balance performance on occluded ReID and holistic ReID tasks. However, recent advances in high-definition surveillance infrastructure and the surge of computational power have significantly improved the applicability of pose-guided methods. In this paper, we propose a multi-granularity feature learning framework designed to fully exploit the rich information within pose cues across three aspects: pose-guided attention, part-based feature extraction, and joint-based semantic feature learning. Specifically, we generate hand-crafted attention maps using multiple pose cues and combine them with learned attention mechanisms. This combined approach directs the network focus toward most non-occluded regions while preventing overemphasis on limited salient areas, thereby enhancing feature diversity. Furthermore, we introduce an improved graph convolution layer to extract joint-based semantic local features, and learn discriminative part-based features through pose-guided body part partition. Extensive experiments validate the network’s superior performance on occluded ReID and demonstrate a favorable balance between occluded ReID and holistic ReID.