Answers are grounded in this paper and cite their sources.
Demonstrates improved stroke classification accuracy in badminton using advanced spatiotemporal models, suggesting enhanced coaching tools.
Automatic stroke classification is the first capability a badminton coaching pipeline needs. It is also hard. The six broadcast stroke classes share most of their kinematic envelope, and the errors that remain are direction-of-motion confusions no single frame can resolve. We treat the problem as a search across the design space of temporal models: a parameter-efficient recurrent encoder with a zero-cost temporal shift, pretrained three-dimensional convolutional backbones, and a temporal-attention pooling head. On a public dataset of 836 clips over six classes, a Kinetics-pretrained 3D residual network reaches 88.7% accuracy across five seeds and 89.2%±2.2 under five-fold cross-validation. A recurrent ResNet-18 encoder manages 71.8%; a parameter-matched video transformer, 83.7%. Replacing the backbone's global average pooling with a temporal-attention head, which weights the spatiotemporal feature sequence before classification, raises accuracy to 91.9%±2.8 and macro-F1 to 0.914. The 3.2-point gain over global pooling is significant under McNemar's test (p=9×10−4), computed from stored per-clip predictions. A two-backbone ensemble reaches 91.0%, and a distilled ResNet-18 student retains 73.4% at one third of the parameters. We report the design search in full, including the interventions that did not help.
No takes yet. Share an insight, caveat, or question.
Kallinatha et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: