Reliability-Conditioned Virtual-View Pose Fusion for Monocular 3D Human Pose Estimation with Real 2D Detector Inputs

Monocular 3D human pose estimation is sensitive to depth ambiguity and upstream 2D-detector error. We investigate a detector-conditioned virtual 2D pose as an auxiliary hypothesis rather than an independent physical measurement. Joint reliability combines training-subject-only monotonic confidence calibration with detector-, joint-, and confidence-specific residual risk before a compact pose lifter. A four-arm factorial design separates observed-only (O), observed plus virtual (OV), observed plus reliability (OR), and observed plus virtual plus reliability (OVR). Across three subject-disjoint splits, three seeds, two RGB detectors, and 48 subject–sequence–camera clusters, OVR improved mean per-joint position error (MPJPE) by 3.049 mm (95% cluster-aware confidence interval (CI): 0.583–4.976); reliability-only improvement was supported, whereas the conditional virtual increment and interaction were not. Validation-only sensitivity covered eight yaw/pitch hypotheses and five multi-view policies; no multi-view policy improved primary MPJPE over the locked −15∘ view. A component-level audit of two synchronized directed camera pairs favored the generator over the observed-pose baseline in all 24 stratified rows, with only two sequence clusters per row. The locked MPI-INF-3DHP Official TS1–TS6 result improved MPJPE by approximately 0.97%, below its prespecified 2% practical threshold, with a CI crossing zero. On an additional external 3DPW cohort (13,579 eligible frames, 24 sequences, YOLO11L-Pose only), OVR improved MPJPE by 1.004 mm (95% CI: −2.965–5.016), while its Procrustes-aligned MPJPE point estimate worsened by 0.640 mm; neither contrast was conclusive. Thus, calibrated reliability is the most consistently supported mechanism, whereas virtual-view dominance, official superiority, and cross-dataset superiority are not established.