An Empirical Study of Open-Set Gesture Recognition on WiFi BVP Features

WiFi-based gesture recognition using domain-independent Body-coordinate Velocity Profile (BVP) features has demonstrated strong cross-domain generalization, but existing methods generally rely on closed-set assumptions, limiting their use in open-world scenarios. In this paper, we present a systematic empirical study of open-set gesture recognition on WiFi BVP features. On the CNN–GRU backbone, we evaluate seven representative OSR methods—MSP, OpenMax, Energy, ODIN, K-NN, prototype, and rejection head—under two backbone strategies (frozen and full fine-tuning), across six unknown gestures and three random seeds. Our experiments show that closed-set Softmax classifiers consistently misclassify unknown gestures with high confidence (49.7–74.3% across six gestures), and that the closed-set high-confidence misclassification rate is strongly associated with open-set AUROC (R2 = 0.918, p = 0.0026), whereas embedding-space geometric distance is not (R2 = 0.018, p = 0.80). This suggests that open-set difficulty may stem from classifier decision-boundary confusion rather than geometric separation. We further find that no single OSR method consistently outperforms all others across unknown gestures, and that frozen and fine-tuned backbones yield comparable open-set performance (typically within ±0.02 AUROC). Finally, we provide multi-objective evaluation and deployment observations, including Pareto frontiers and multiple operating points. These results offer a systematic benchmark for open-set WiFi gesture recognition and highlight the importance of classifier-level factors in determining open-set difficulty.