| 方法 | Backbones | Top1 (%) | GFLOPs |
| LiteEval [24] | MobileNetV2 + ResNet101 | 61.0 | 99.0 |
| SCSampler [12] | MobileNetV2 + ResNet50 | 70.8 | 42.0 |
| AR-Net [14] | MobileNetV2 + ResNet50 | 71.7 | 32.0 |
| D-STEP [26] | ResNet50 | 65.4 | 12.4 |
| 本文方法 | TimeSformer | 73.9 | 25.9 |