Label-efficient segmentation of high-frame-rate table tennis video supports scene-level process indicators

Camera-based video can provide a low-intrusion sensing source for sports movement analysis, but dense pixel-level annotation remains costly when semantic masks are needed for downstream measurement. We present EduSeg, a label-efficient semi-supervised video semantic segmentation framework for high-frame-rate table tennis videos. At a 20% label budget, EduSeg achieves 0.750 ± 0.003 foreground mIoU. Under a stricter temporally blockwise protocol that eliminates labeled–unlabeled raw-frame overlap, EduSeg-base achieves 0.735 ± 0.004 mIoU, compared with 0.720 ± 0.002 for a video-specific TDC baseline, 0.709 ± 0.004 for UniMatch, and 0.686 ± 0.004 for a temporal supervised baseline. Across five held-out-video folds within OpenTTGames, EduSeg-base achieves 0.720 ± 0.011 mIoU and retains the highest mean in each fold. Sensitivity analyses place nine frames on an accuracy–stability–cost plateau, while controlled degradation tests show progressive performance loss as corruption severity increases. Controlled mask perturbations establish a descriptor-level error budget for the Process Analytics Layer, with table-anchor errors producing the largest cross-indicator propagation. Expert-rating associations remain stable under bootstrap and leave-one-out checks, although re-centering retains the widest uncertainty.