CaM-PDA: Confidence-Aware Prior Depth Completion with a Mixture of Experts for Low-Cost, Spray-Free 3D Reconstruction of Reflective Low-Pressure Compressor Blades

Reflective low-pressure compressor blades produce missing and biased returns in low-cost depth cameras. CaM-PDA combines geometric anchor screening with prior-guided depth completion and residual experts for reflection, surface variation and boundaries. Balanced screening retains reliable observations for metric alignment, while a pretrained general path remains active alongside the specialist residuals. On 11,520 DREDS-CatNovel targets and 654 NYUv2 images, CaM-PDA achieves whole-image AbsRel of 0.015880 and 0.023305, reducing error against native PDA by 49.98% and 57.61%, respectively. Within annotated specular/transparent DREDS regions, AbsRel decreases by 22.31% against native PDA. Comparisons with other native methods retain dataset-dependent strengths. Adaptation on 854 unique frames yields modest, region-dependent expert refinement, assessed separately through component ablations and three-seed matched controls. A blade reconstruction compared with blue-light geometry has directional mean surface distances of 3.514 and 3.042 mm after rigid alignment. These results support spray-free completion of visible blade geometry from affordable RGB-D observations.