TY - GEN
T1 - Int&Int
T2 - 18th IEEE Conference on Industrial Electronics and Applications, ICIEA 2023
AU - Qi, Xiangyuan
AU - He, Zhen
AU - Wang, Qiang
N1 - Publisher Copyright:
© 2023 IEEE.
PY - 2023
Y1 - 2023
N2 - Human action recognition has recently attracted a lot of researchers, while most of them mainly focus on the local dynamics or global information alone and cannot focus on both the intensity and the integrity of the action. In this work, we propose a two-pathway Int&Int network(Intensity&Integrity) for skeleton-based action recognition to satisfy both aspects, where the great complementarity between the two pathways further enhances the performance. Besides, for Integrity pathway, we apply the uniform sampling strategy. For Intensity pathway, we introduce the intensity-dependent sampling, where a clip composed of consecutive frames around the frame with the largest motion intensity is sampled. Moreover, we explain various definitions of the motion intensity containing different semantic information based on the extracted 2D human poses. For each pathway, the poses are represented by a 3D heatmap volume and 3D-CNNs of both pathways have the same architecture. Late fusion is used to ensemble them. The model is evaluated on two action recognition datasets, FineGYM-99 and HMDB-51, and it achieves superior performance on both of them. The code has been shown at https://github.com/SarahQi666/Int-and-Int.
AB - Human action recognition has recently attracted a lot of researchers, while most of them mainly focus on the local dynamics or global information alone and cannot focus on both the intensity and the integrity of the action. In this work, we propose a two-pathway Int&Int network(Intensity&Integrity) for skeleton-based action recognition to satisfy both aspects, where the great complementarity between the two pathways further enhances the performance. Besides, for Integrity pathway, we apply the uniform sampling strategy. For Intensity pathway, we introduce the intensity-dependent sampling, where a clip composed of consecutive frames around the frame with the largest motion intensity is sampled. Moreover, we explain various definitions of the motion intensity containing different semantic information based on the extracted 2D human poses. For each pathway, the poses are represented by a 3D heatmap volume and 3D-CNNs of both pathways have the same architecture. Late fusion is used to ensemble them. The model is evaluated on two action recognition datasets, FineGYM-99 and HMDB-51, and it achieves superior performance on both of them. The code has been shown at https://github.com/SarahQi666/Int-and-Int.
KW - Intensity-dependent sampling
KW - Skeleton-based action recognition
KW - Two-pathway network
UR - https://www.scopus.com/pages/publications/85173603204
U2 - 10.1109/ICIEA58696.2023.10241867
DO - 10.1109/ICIEA58696.2023.10241867
M3 - 会议稿件
AN - SCOPUS:85173603204
T3 - Proceedings of the 18th IEEE Conference on Industrial Electronics and Applications, ICIEA 2023
SP - 1477
EP - 1482
BT - Proceedings of the 18th IEEE Conference on Industrial Electronics and Applications, ICIEA 2023
A2 - Cai, Wenjian
A2 - Yang, Guilin
A2 - Qiu, Jun
A2 - Gao, Tingting
A2 - Jiang, Lijun
A2 - Zheng, Tianjiang
A2 - Wang, Xinli
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 18 August 2023 through 22 August 2023
ER -