• Laser & Optoelectronics Progress
  • Vol. 56, Issue 12, 121004 (2019)
Ningxiao Li, Guodong Wang*, Yanjie Wang, Shiyu Hu, and Liangliang Wang
Author Affiliations
  • College of Computer Science & Technology, Qingdao University, Qingdao, Shandong 266071, China
  • show less
    DOI: 10.3788/LOP56.121004 Cite this Article Set citation alerts
    Ningxiao Li, Guodong Wang, Yanjie Wang, Shiyu Hu, Liangliang Wang. Video Classification Based on Three-Dimensional Squeeze Excitation Module[J]. Laser & Optoelectronics Progress, 2019, 56(12): 121004 Copy Citation Text show less
    Unit structure of module. (a) 2D ResNet unit; (b) 3D ResNet unit; (c) 3D SE_FC ResNet unit
    Fig. 1. Unit structure of module. (a) 2D ResNet unit; (b) 3D ResNet unit; (c) 3D SE_FC ResNet unit
    Flow chart of feature extraction
    Fig. 2. Flow chart of feature extraction
    Loss at first stage of UCF101 training
    Fig. 3. Loss at first stage of UCF101 training
    Loss at second stage of UCF101 training
    Fig. 4. Loss at second stage of UCF101 training
    Loss of HMDB51 training
    Fig. 5. Loss of HMDB51 training
    MethodPretrainingdatasetAverage validationaccuracy /%
    Top1Top5
    3D ResNet5043.867.3
    3D ResNet10143.468.8
    3D SE_FC ResNet50,T=1No42.566.1
    3D SE_FC ResNet50,T=444.969.6
    3D SE_FC ResNet50,T=845.568.2
    Table 1. Average validation accuracy comparison on UCF101 dataset
    MethodPretrainingdatasetAverage validationaccuracy /%
    Top1Top5
    3D ResNet5015.643.0
    3D ResNet10115.241.0
    3D SE_FC ResNet50,T=1UCF10116.143.6
    3D SE_FC ResNet50,T=418.746.2
    3D SE_FC ResNet50,T=817.744.6
    Table 2. Average validation accuracy comparison on HMDB51 dataset
    MethodTime /ms
    3D ResNet50104
    3D ResNet101111
    3D SE_FC ResNet50,T=1112
    3D SE_FC ResNet50,T=4110
    3D SE_FC ResNet50,T=8114
    Table 3. Average time comparison for each video classification
    MethodPretrainingdatasetTest accuracy /%
    HMDB51UCF101
    3D ResNet18Kinetics56.484.4
    3D ResNet3459.187.7
    3D ResNet5061.089.3
    3D ResNet10161.788.9
    3D ResNet20063.589.6
    3D DenseNet-12159.687.6
    Method in Ref. [21]-59.488.0
    Method in Ref.[22]--88.6
    Method in Ref. [23]--85.9
    Method in Ref. [24]--88.6
    Method in Ref.[25]sports 1M-82.3
    3D SE_FC ResNet50,T=1(ours)Kinetics61.389.0
    3D SE_FC ResNet50,T=4(ours)59.690.1
    3D SE_FC ResNet50,T=8(ours)59.089.5
    Table 4. Test accuracy comparison on HMDB51 and UCF101 datasets
    Ningxiao Li, Guodong Wang, Yanjie Wang, Shiyu Hu, Liangliang Wang. Video Classification Based on Three-Dimensional Squeeze Excitation Module[J]. Laser & Optoelectronics Progress, 2019, 56(12): 121004
    Download Citation