Integrating Temporal and Spatial Attention for Video Action Recognition

Table 2

Comparisons with other methods on UCF101 dataset.

Model	Pretraining dataset	Accuracy (%)	GFLOPs
C3D [20]	Sports-1M	82.3	38.57
TRN [21]	—	83.5	83.83
Res3D [22]	Sports-1M	85.8	—
P3D [23]	Imagenet + Sports-1M	88.6	18.51
T3D [24]	Kinetics-400	90.3	—
TSN [8]	Imagenet + Kinetics-400	91.1	80
R(2 + 1)D [25]	Sports-1M	93.6	41.69
TSM [26]	Kinetics-400	95.5	32.88
I3D RGB [27]	Imagenet + Kinetics-400	95.6	108
T-CNN [12]	Kinetics-400	95.3	15.78
T-CNN + spatial	Kinetics-400	96.7	52.3