方法

Backbones

mAP (%)

GFLOPs

FrameGlimpses [8]

VGG

60.2

32.9

AdaFrame [9]

ResNet101

71.5

79.0

LiteEval [24]

MobileNetV2+ResNet101

72.7

95.1

ListenToLook [10]

MobileNetV2+ResNet50

72.3

81.4

SCSampler [12]

MobileNetV2+ResNet50

72.9

42.0

AR-Net [14]

MobileNetV2+ResNet50

73.8

33.5

FrameExit [11]

ResNet50

76.1

26.1

TSQNet [25]

ResNet50

76.6

26.1

本文方法

TimeSformer

77.8

25.9