BROWSE
Detail
Conference Paper
Global Positional Self-Attention for Skeleton-Based Action Recognition
Cited 1 time in 
Share 
Share
- Authors
- Jaehwan Kim, Junsuk Lee
- Issue Date
- 2022-08
- Citation
- International Conference on Pattern Recognition (ICPR) 2022, pp.1-7
- Language
- English
- Type
- Conference Paper
- Abstract
- In this paper, we introduce a novel global positional self-attention network, which represents the spatially structured dependencies and the globally ordered semantic information for skeleton-based action recognition. The structured dependencies are learned through our spatial self-attention composed of the weighted sum of three correlations1, each of which is a correlation between spatial elements, between sequential positions, and between structural geodesic positions. In addition, our channel self-attention with the global average pooling and channel-wise positional encoding operations allows capturing the globally ordered semantic and positional dependencies between channels. The self-attention modules are incorporated into standard CNNs, which is referred to as global positional self-attention network, GPS-Net. Our GPS-Net is a simple yet effective network, which leads to make accurate action predictions. We evaluate our GPS-Net on open large-scaled benchmark datasets NTU-RGB+D and DHG for 3D body action and hand gesture recognitions, respectively. Moreover, our self-converted NTU dataset for compatibility with OpenPose skeletons is used for the experiments. Numerical and visual comparisons with the existing state-of-the-art methods confirm the usefulness of the proposed network.
- KSP Keywords
- Action recognition, Benchmark datasets, Hand Gesture, existing state, semantic information, state-of-The-Art, weighted sum