Table 4 Comparison of recognition accuracy with state-of-the-art methods on HMDB51 and UCF101 datasets
From: Fused behavior recognition model based on attention mechanism
Methods | Input modality | Pre_training | HMDB51 (%) | UCF101 (%) |
|---|---|---|---|---|
HOG/HOF [1] | RGB | – | 20.44 | – |
IDT [17] | RGB | – | 57.2 | 85.9 |
MIFS [18] | RGB | – | 65.1 | 89.1 |
ECO-Lite (16 frames) [12] | RGB | Kinetics | 68.2 | 91.6 |
ECO (16 frames) [12] | RGB | Kinetics | 68.5 | 92.8 |
ResNext-101 [19] | RGB | Kinetics | 63.8 | 90.7 |
Res3D [15] | RGB | Sports-1 M | 54.9 | 85.8 |
I3D [9] | RGB | Kinetics | 74.5 | 95.4 |
ResNet101 [19] | RGB | Kinetics | 61.7 | 88.9 |
DTTP (split 1) [20] | RGB | ImageNet | 61.5 | 89.7 |
RSN [21] | RGB | – | 55.9 | 87.5 |
Two-stream (fusion by SVM) [5] | RGB, Optical flow | ILSVRC | 59.4 | 88.0 |
VGG16 + TSN [22] | RGB,Optical flow | ImageNet | 67.3 | 92.1 |
ResNet34-3DRes18 (16 frames) | RGB | Kinetics | 70.997 | 92.143 |
Res34-SE-IM-Net (16 frames) | RGB | Kinetics | 71.85 | 92.196 |