Skip to main content
. 2020 Feb 17;20(4):1085. doi: 10.3390/s20041085

Table 6.

Comparison of recognition accuracy with different network parameter initialization methods. I3D: inflated 3D convnet.

Methods Pre-Training Spatial ConvNets Temporal ConvNets Two-Stream
I3D 93.30% 94.68% 95.45%
I3D ImageNet + Kinetics 97.89% 98.28% 98.99%