PSEG improves human pose estimation via plug-and-play structural edge guidance

Abstract

Human pose estimation in dense occlusion suffers from keypoint misdetection and joint misalignment. Existing methods struggle to exploit structural edge constraints owing to the lack of high-quality human edge data and dedicated detectors. We propose PSEG (Plug-and-Play Structural Edge Guidance), a framework that addresses occlusion in crowded multi-person pose estimation. We introduce PSEG-Bench, the first large-scale 2D multi-person edge dataset generated via SMPL projection with APDC-RPDC adaptive morphological optimization, providing reliable supervision. We develop a lightweight PSEG module trained by semi-supervised learning using 200,000 real unlabeled frames to bridge the synthetic-to-real domain gap between SMPL-projected and real edges, with dynamic confidence thresholds. Ablation studies verify that our input-end single-scale injection adds only 2.39% parameters while outperforming multi-scale injection; the loss weight α = 1.1 optimally balances edge-background class imbalance, while reward-driven and gradient consistency constraints suppress Canny-induced noise. Strict non-overlapping dataset partitioning eliminates data leakage. Under motion blur (σ = 5) and low-light (γ = 0.5) degradation, PSEG delivers up to 6.4 AP recovery. On OCHuman, it yields 1.2–2.1% absolute AP gain. On CrowdPose, PSEG improves HRNet and TransPose by 0.8–2.6 AP (average 1.52). On MSCOCO, it boosts mAP by 1.1–2.5 (average 1.62). Experiments show that PSEG suppresses false activations and accelerates convergence. Code: https://github.com/wilson-up1/PSEG-main .