RGB-D Human Action Recognition of Deep Feature Enhancement and Fusion Using Two-Stream ConvNet

Table 3

Comparison of accuracy of adding nonlocal to different locations of st-gcn.

Network	Top 1	Top 5
Baseline	81.5%
1-block	85.23%	98.74%
2-block	86.43%	98.62%
3-block	85.43%	97.12%
4-block	82.14%	97.2%
5-block	85.55%	97.31%
1-2-block	85.63%	96.32%
1-3-block	84.08%	95.67%
1-4-block	84.24%	92.35%
2-2-block	87.62%	97.3%
2-3-block	84.1%	95.2%
2-4-block	84.41%	94.69%
3-3-block	83.77%	94.12%
3-4-block	80.19%	91.63%
4-4-block	77.09%	91.03%
5-5-block	77.75%	90.12%