00 本有行 .10 190 2 540 K Value of KNN (a)DSW-INN vs DTW-INN. (b)DSW-kNN vs DTW-kNN. (c)Methods on one trainer's data. (d)Methods on four trainers'data. Figure 6.Generalization capabilities of different supervised learning methods 50 so SVM (10-3ccmSVM-406m) A BC D G Gesture Number of Fine-tuning Samples per Gesture Gesture Label Gesture Label (a)Different angles. (b)Different distances. (a)DSW-based AP clustering. (b)DTW-based AP clustering. Figure 7.Robusinesstunin of DSW.INN. Figure 8.Clustering results based on DSW or DTW. VALIDITY INDEX OF CLUSTERING Index DSW DTW on the geometric relationship to find the cluster center is not 0.9772 0.4380 applicable for these two distances,as they do not satisfy the FMI 0.9884 0.6094 triangle inequality.Thus,we use the Affinity Propagation(AP) RI 0.9974 0.9110 clustering algorithm [37].We adjust the damping coefficient generalization accuracy of DSW-1NN,DTW-INN,CNN and and preference factor to achieve aggregation of different num- SVM is 99.33%,91.11%,96.61%and 95.70%,respectively. bers of clusters.As we have nine gesture types,we choose to DSW-INN is also robust to gestures at different angles and converge at nine clusters.As shown in Figure 8(a),the distance different distances.To evaluate the influence of different angles defined by DSW correctly cluster the gesture of the same type on the model,we request a volunteer to perform gestures at into the same category.However,the DTW-based clustering three angles (0,45,and 90 degree with respect to the center does not give the correct clustering results.We further use of the phone)while maintaining an equal distance to the three common external indicators to represent cluster validity, phone.We use the 0-degree gesture samples as the training such as Jaccard Coefficient (JC),Fowles and Mallows Index set,which has five samples for each gesture type.Then,we (FMI)and Rand Index (RD).where a larger indicator means evaluate the performance on test datasets at three different better clustering performance.As shown in Table II.DSW- angles,each contains 225 gesture samples.After 50 rounds based clustering has higher scores for all three indicators than of Monte Carlo cross-validation,we find that the accuracy of DTW.This proves that the similarity measure defined by DSW DSW-INN at different angles is 100%.98.85%,and 99.89%. can be used for clustering large-scale unlabeled data and can respectively.We also request the volunteer to perform gestures even guide the design of gestures. at three different distances (10~20 cm,20~30 cm and 30~40 VI.CONCLUSION cm)from the mobile phone.We use the 45 samples in the In this paper,we introduce DSW,a dynamical speed ad- 20~30 cm region as the source domain and DSW-INN has an aptive matching scheme for gesture signals.DSW rescales the accuracy of 99.86%and 91.64%on the testing sets of 10~20 speed distributions of gesture samples while keeping track of cm and 30~40 cm.This means that the DSW-INN algorithm the movement distance at different stages.Our experimental is robust to distance changes within 30 cm.When the distance results show that DSW provides a robust similarity measure changes by more than 30 cm,the accuracy of the DSW-INN between gesture samples and can work for both one-shot algorithm is reduced to 91%due to the signal attenuation. learning in supervised gesture recognition and unsupervised However,we can fine-tune the model by adding a small learning tasks.While we mainly focus on sound-based signals number of samples in the target domain without retraining.As and STFT features,we believe that DSW can be readily shown in Figure 7(b),after adding five target-domain samples applied to Wi-Fi based gesture recognition systems and other for each gesture type,the accuracy of the 30~40 cm region time-frequency analysis schemes,such as DWT or HHT. is increased to 98.61%.In the same situation,CNN and SVM VII.ACKNOWLEDGEMENT can only reach 79.8%and 73.57%before fine-tuning,and the We would like to thank our anonymous shepherd and model parameters need to be retrained for the target domain. reviewers for their valuable comments.This work is partially C.Evaluation for Clustering Tasks supported by National Natural Science Foundation of China Clustering is another application scenario for similarity under Numbers 61872173,61972192,and Collaborative In- measure like DSW and DTW.The clustering algorithm based novation Center of Novel Software Technology.1 2 3 4 5 6 7 8 9 10 Number of Training Samples per Gesture 70 80 90 100 Accuracy(%) 1-DSW 2-DSW 3-DSW 4-DSW 1-DTW 2-DTW 3-DTW 4-DTW (a) DSW-1NN vs DTW-1NN. 1357 K Value of KNN 80 85 90 95 100 Accuracy(%) 1-DSW 2-DSW 3-DSW 4-DSW 1-DTW 2-DTW 3-DTW 4-DTW (b) DSW-kNN vs DTW-kNN. 0 45 90 135 180 225 Number of Training Samples 50 60 70 80 90 100 Accuracy(%) DSW DTW SVM CNN (c) Methods on one trainer’s data. 0 180 360 540 720 900 Number of Training Samples 75 80 85 90 95 100 Accuracy(%) DSW DTW SVM CNN (d) Methods on four trainers’ data. Figure 6. Generalization capabilities of different supervised learning methods. ABCDEFGH I Gesture 90 92 94 96 98 100 Accuracy(%) 0 Degree (Baseline) 45 Degree 90 Degree (a) Different angles. 012345 Number of Fine-tuning Samples per Gesture 70 75 80 85 90 95 100 Accuracy(%) DSW (10~20cm) CNN (10~20cm) SVM (10~20cm) DSW (30~40cm) CNN (30~40cm) SVM (30~40cm) (b) Different distances. Figure 7. Robustness and fine-tuning of DSW-1NN. Table II VALIDITY INDEX OF CLUSTERING Index DSW DTW JC 0.9772 0.4380 FMI 0.9884 0.6094 RI 0.9974 0.9110 generalization accuracy of DSW-1NN, DTW-1NN, CNN and SVM is 99.33%, 91.11%, 96.61% and 95.70%, respectively. DSW-1NN is also robust to gestures at different angles and different distances. To evaluate the influence of different angles on the model, we request a volunteer to perform gestures at three angles (0, 45, and 90 degree with respect to the center of the phone) while maintaining an equal distance to the phone. We use the 0-degree gesture samples as the training set, which has five samples for each gesture type. Then, we evaluate the performance on test datasets at three different angles, each contains 225 gesture samples. After 50 rounds of Monte Carlo cross-validation, we find that the accuracy of DSW-1NN at different angles is 100%, 98.85%, and 99.89%, respectively. We also request the volunteer to perform gestures at three different distances (10⇠20 cm, 20⇠30 cm and 30⇠40 cm) from the mobile phone. We use the 45 samples in the 20⇠30 cm region as the source domain and DSW-1NN has an accuracy of 99.86% and 91.64% on the testing sets of 10⇠20 cm and 30⇠40 cm. This means that the DSW-1NN algorithm is robust to distance changes within 30 cm. When the distance changes by more than 30 cm, the accuracy of the DSW-1NN algorithm is reduced to 91% due to the signal attenuation. However, we can fine-tune the model by adding a small number of samples in the target domain without retraining. As shown in Figure 7(b), after adding five target-domain samples for each gesture type, the accuracy of the 30⇠40 cm region is increased to 98.61%. In the same situation, CNN and SVM can only reach 79.8% and 73.57% before fine-tuning, and the model parameters need to be retrained for the target domain. C. Evaluation for Clustering Tasks Clustering is another application scenario for similarity measure like DSW and DTW. The clustering algorithm based ABCDEFGH I Gesture Label 0 50 100 150 200 250 Number of Samples 0 1 2 3 4 5 6 7 8 (a) DSW-based AP clustering. ABCDEFGH I Gesture Label 0 50 100 150 200 250 Number of Samples 0 1 2 3 4 5 6 7 8 (b) DTW-based AP clustering. Figure 8. Clustering results based on DSW or DTW. on the geometric relationship to find the cluster center is not applicable for these two distances, as they do not satisfy the triangle inequality. Thus, we use the Affinity Propagation (AP) clustering algorithm [37]. We adjust the damping coefficient and preference factor to achieve aggregation of different num￾bers of clusters. As we have nine gesture types, we choose to converge at nine clusters. As shown in Figure 8(a), the distance defined by DSW correctly cluster the gesture of the same type into the same category. However, the DTW-based clustering does not give the correct clustering results. We further use three common external indicators to represent cluster validity, such as Jaccard Coefficient (JC), Fowles and Mallows Index (FMI) and Rand Index (RI), where a larger indicator means better clustering performance. As shown in Table II, DSW￾based clustering has higher scores for all three indicators than DTW. This proves that the similarity measure defined by DSW can be used for clustering large-scale unlabeled data and can even guide the design of gestures. VI. CONCLUSION In this paper, we introduce DSW, a dynamical speed ad￾aptive matching scheme for gesture signals. DSW rescales the speed distributions of gesture samples while keeping track of the movement distance at different stages. Our experimental results show that DSW provides a robust similarity measure between gesture samples and can work for both one-shot learning in supervised gesture recognition and unsupervised learning tasks. While we mainly focus on sound-based signals and STFT features, we believe that DSW can be readily applied to Wi-Fi based gesture recognition systems and other time-frequency analysis schemes, such as DWT or HHT. VII. ACKNOWLEDGEMENT We would like to thank our anonymous shepherd and reviewers for their valuable comments. This work is partially supported by National Natural Science Foundation of China under Numbers 61872173 , 61972192, and Collaborative In￾novation Center of Novel Software Technology