Third-view industrial assembly videos exhibit long transitional periods, weak hand–tool cues under occlusion, and strict efficiency requirements, which jointly challenge existing video action recognition models. This study proposes an efficiency-oriented video transformer framework that explicitly targets temporal redundancy while preserving fine-grained interaction evidence. A motion-guided PreFocusSelector performs input-level frame selection and outputs an effective subsequence for backbone inference, thereby reducing redundant spatiotemporal token processing. To compensate for subtle local dynamics, an AxisSplit Adapter is introduced to enhance short-range spatiotemporal modeling with lightweight overhead. In addition, a temporal attention pooling module replaces uniform temporal averaging by learning discriminative temporal weights, improving video-level evidence integration under sparse signals. Experiments are conducted on a third-view assembly action dataset filtered to 1759 clips with a consistent train/validation/test split. Under a fixed evaluation protocol and 10 independent runs, the proposed method achieves 96.41% ± 0.22% Top-1 accuracy and 94.83% ± 0.26% Macro-F1, while reducing computational cost from 88.1 to 37.6 G FLOPs with modest parameter increase. A common frame subsampling baseline is included to contextualize efficiency comparisons, and comprehensive implementation details and ablation studies are provided to support reproducibility.
Zhang et al. (Wed,) studied this question.