Integrating Temporal and Spatial Attention for Video Action Recognition

Table 3

Comparisons with other methods on HMDB51 dataset.

Model	Pretraining dataset	Accuracy (%)	GFLOPs
Res3D [22]	Sports-1M	54.9	—
T3D [24]	Kinetics-400	59.2	—
R(2 + 1)D [25]	Sports-1M	66.6	41.69
TSM [26]	Kinetics-400	73.6	32.88
I3D RGB [27]	Imagenet + Kinetics-400	74.8	108
T-CNN [12]	Kinetics-400	73.3	15.78
T-CNN + spatial	Kinetics-400	75.2	52.3