Transferable Feature Representation for Visible-to-Infrared Cross-Dataset Human Action Recognition

Table 2

The average recognition precision (%) of No Adaptation (NA), KEMA, and the proposed CDFAG in target dataset InfAR. S_train and T_train denote the number of training samples in source and target datasets, respectively. Best accuracies of each column for each method are marked in italics.

Settings	T_train = 25	T_train = 20	T_train = 15	T_train = 10	T_train = 5
S_train = 45
NA
KEMA
CDFAG
S_train = 40
NA
KEMA	69.67 ± 1.43
CDFAG	75.42 ± 3.13			62.42 ± 2.59
S_train = 35
NA
KEMA			62.67 ± 2.76	56.92 ± 2.74
CDFAG
S_train = 30
NA
KEMA
CDFAG
S_train = 25
NA
KEMA
CDFAG		71.58 ± 2.42	68.08 ± 3.87		50.75 ± 2.19
S_train = 20
NA
KEMA					46.25 ± 2.38
CDFAG
S_train = 15
NA			54.17 ± 4.31		33.83 ± 3.22
KEMA		67.08 ± 3.27
CDFAG
S_train = 10
NA	66.50 ± 2.01	59.75 ± 3.14		46.75 ± 3.53
KEMA
CDFAG
S_train = 5
NA
KEMA
CDFAG