Research Article

Continuous Gesture Sequences Recognition Based on Few-Shot Learning

Table 5

Ablation study of the gesture feature extractor model.

ModelSimilarity measureRWTH
5-way
1-shot acc
RWTH
5-way
3-shot acc
SMD
4-way
1-shot acc
SMD
4-way
3-shot acc

FullMSE79.45%80.72%91.14%91.11%
Cosine89.73%90.16%92.17%93.23%

None MSE79.41%80.47%90.36%91.34%
Cosine85.62%85.95%91.69%92.10%

None MSE75.71%75.64%89.97%89.84%
Cosine82.88%84.13%90.38%90.75%

None MSE75.34%77.35%86.79%88.19%
Cosine84.25%85.13%90.20%91.73%

None MSE79.42%78.68%77.69%77.42%
Cosine88.36%89.25%81.99%83.22%

None MSE72.60%74.26%76.40%77.46%
Cosine85.62%85.45%80.40%82.31%

None MSE74.66%74.24%81.18%82.54%
Cosine86.30%87.83%86.27%87.22%

None
SA module
MSE76.36%77.69%88.18%88.78%
Cosine88.54%88.83%90.97%91.02%