Answers are grounded in this paper and cite their sources.
A critical challenge in skeleton-based action recognition lies in effectively extracting meaningful features from skeleton joints. However, existing state-of-the-art networks for this task are often complex and heavily parameterized, leading to inefficient training and inference times, particularly on largescale datasets. In this work, we propose a simple yet efficient baseline for skeleton-based Human Action Recognition (HAR), building on our previous model, ImpSGN. Our new approach features a lighter and faster architecture while preserving the same performance capabilities. The proposed architecture employs attention-based graph convolution to effectively capture the intricate interconnections within skeletal structures. We report competitive accuracy on the large-scale NTU-RGB+D 60 dataset, achieving $90.0 %$ and $95.0 %$ on the Cross-Subject and CrossView benchmarks, respectively. On NTU-RGB+D 120, the model achieves $84.7 %$ and $85.8 %$ on the Cross-Subject and Cross-Setup benchmarks, respectively. This work demonstrates a significant improvement over our previous model, ImpSGN, by extracting more discriminative spatial and temporal features, all while reducing the number of parameters by $60 %$.
No takes yet. Share an insight, caveat, or question.
Mansouri et al. (2025) studied this question.