Keyframe Extraction Algorithm for Continuous Sign-Language Videos Using Angular Displacement and Sequence Check Metrics

Table 4

A comparison of the WER metrics obtained by different systems on sign language recognition tasks.

Methods	dev	Test
Deep sign [47]	38.3	38.8
SubUNets [48]	40.8	40.7
SF-Net [49]	35.6	34.9
SAN [45]	29	29.7
SAN [45] + Zernike’s moment [35]	38.1	38.2
SAN [45] + pixel difference [13]	40.1	40.2
SAN [45] + gradient based [27]	42.7	43.2
SAN + FSC2 (proposed)	28.6	28.8
VAC [46]	21.2	22.3
VAC [46] + Zernike’s moment [35]	28.1	28.2
VAC [46] + pixel difference [13]	32.1	33.2
VAC [46] + gradient based [27]	31.7	31.8
VAC + FSC2 (proposed)	20.8	21.9

Tested on the RWTH-PHOENIX-weather 2014 dataset. A lower WER value is better. The highlighted value suggests that FSC2 performs well when integrated with existing systems.