Query-Guided Refinement and Separation for Crowded Human Instance Segmentation

Crowded human instance segmentation is challenged by small persons, partial occlusion, close inter-person contact, and boundary adhesion, which often cause missed detections, incomplete masks, and merged instances. This paper proposes QHIS-Net (Query-Guided Human Instance Separation Network), a query-guided refinement-and-separation framework that jointly addresses intra-query local mask degradation and inter-query response conflict. The Occlusion-Aware Human-Mask Refinement Module (OHMRM) uses boundary- and uncertainty-aware weighting to focus query-conditioned residual correction on locally degraded human-mask regions, improving instance recovery and mask completeness rather than acting only as a contour-sharpening module. The Adjacent-Instance Separation Constraint Module (AISCM) complements this recovery mechanism by selectively penalizing excessive soft overlap among high-confidence human queries during training, without adding an inference-time branch. CIHP parsing annotations are converted into COCO-style person-instance masks for evaluation, and cross-dataset generalization is tested on OCHuman without fine-tuning. On CIHP-val, QHIS-Net improves AP50:95 from 0.764 to 0.774, AP75 from 0.841 to 0.850, and APs from 0.186 to 0.212 over M2FP. On OCHuman, AP50:95 improves from 0.416 to 0.425. Targeted analyses further show improved touching-instance separation, instance recovery, and a more evident advantage in the highest-density subset.