Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

TSGCNeXt: Dynamic-Static Multi-Graph Convolution for Efficient Skeleton-Based Action Recognition with Long-term Learning Potential

Created by
  • Haebom

Author

Dongjingdin Liu, Pengpeng Chen, Miao Yao, Yijing Lu, Zijie Cai, Yuxin Tian

Outline

This paper proposes the Temporal-Spatio Graph ConvNeXt (TSGCNeXt) model to improve the efficiency of skeletal-based action recognition in long-term time-series data. TSGCNeXt utilizes a novel, simple-to-structure graph learning mechanism, Dynamic-Static Separate Multi-graph Convolution (DS-SMG), to aggregate features from multiple independent topological graphs and prevent node information loss during dynamic convolution. Furthermore, it improves the backpropagation computation of dynamic graph learning by 55.08% through a graph convolution learning acceleration mechanism, and efficiently models long-term time-series features using three spatiotemporal learning modules. It outperforms existing methods on large-scale datasets, NTU RGB+D 60 and 120, and achieves state-of-the-art performance by introducing an EMA model for multi-stream fusion. It achieves accuracies of 90.22% and 91.74% on the cross-subject and cross-set datasets of NTU 120, respectively.

Takeaways, Limitations

•
Takeaways:
◦
We present an efficient learning mechanism for long-term time series skeleton data.
◦
We propose a novel graph convolutional structure that reduces the complexity and improves the speed of existing GCNs.
◦
We demonstrate the superiority of our model by achieving state-of-the-art performance on large datasets.
◦
Reduces computational complexity through efficient dynamic graph learning and acceleration mechanisms.
•
Limitations:
◦
Further research is needed to evaluate the generalization performance of the proposed method.
◦
Research is needed on fusion with other modalities (e.g. RGB images).
◦
It may be optimized for a specific dataset, and its performance on other datasets requires further validation.
👍
Made with Slashpage