The Multidimensional Motion Features of Spatial Depth Feature Maps: An Effective Motion Information Representation Method for Video-Based Action Recognition

Table 2

Comparison of our results to the state-of-the-arts on action recognition datasets UCF101 and HMDB51.

Method	UCF101 (%)	HMDB51 (%)
LRCN [13]	82.9
Dense trajectories [42]	84.2
Composite LSTM model [33]	84.3
Soft attention [34]	84.9	39.87
Two-stream ConvNet [4]	88.0	59.4
C3D [2]	90.4	—
TDD [43]	91.5	65.9
Action transformations [44]	92.4	62.0
TSN [23]	94	68.5
DOVF [21]	94.9	71.7
TLE [22]	95.6	71.1
Ours (MDF:MD3DCNN+4)	96.2	`65.5