Eliminating False Positives in Single-Human Parsing via Prompt Filtering and Pose-Aware Segmentation

While existing single human parsing (SHP) models achieve strong performance on standard benchmarks, they can still erroneously predict segments for parts not actually present, particularly when faced with unseen or culturally unique garments. To tackle this limitation, we propose a novel parsing approach that integrates Grounding DINO, the Segment Anything Model (SAM), and a lightweight pose estimator (DWPose) into a unified pipeline. Additionally, we introduce a new challenging dataset comprising 200 carefully annotated images of Vietnamese high school students wearing traditional Ao Dai uniforms, aimed at evaluating model robustness on culturally specific attire. By combining prompt-based semantic filtering and pose-aware segmentation, our method effectively suppresses false-positive predictions and accurately recovers occluded or partially visible body parts. Experimental evaluations demonstrate that our approach achieves a mean Intersection-over-Union (mIoU) of 94.6% on the Ao Dai dataset, significantly surpassing existing baseline methods by significantly reducing hallucinated segments. Moreover, the proposed pipeline maintains competitive performance on the standard ATR dataset, reaching an mIoU of 85.0%. These results underscore the importance of prompt-driven segmentation combined with pose-aware refinement, setting a new benchmark for reliable and culturally adaptive human parsing systems.