Title: RETIT: a recursive symmetry temporal interactive transformer used for compressed video action recognition

Authors: Jiyuan Wang; Huilan Luo; Chanjuan Wang; Ting Li

Addresses: Jiangxi Province Key Laboratory of Multidimensional Intelligent Perception and Control, Jiangxi University of Science and Technology, Ganzhou, 341000, China ' Jiangxi Province Key Laboratory of Multidimensional Intelligent Perception and Control, Jiangxi University of Science and Technology, Ganzhou, 341000, China ' Jiangxi Province Key Laboratory of Multidimensional Intelligent Perception and Control, Jiangxi University of Science and Technology, Ganzhou, 341000, China ' Jiangxi Province Key Laboratory of Multidimensional Intelligent Perception and Control, Jiangxi University of Science and Technology, Ganzhou, 341000, China

Abstract: The growing adoption of compressed video across diverse applications underscores the demand for efficient action recognition methods. Traditional RGB-based methods face limitations, especially because they depend heavily on computationally intensive optical flow for temporal analysis. We introduce the recursive symmetric temporal interaction transformer (RETIT), which leverages the transformer architecture to enhance global temporal interactions directly from compressed video data. RETIT employs recursive strategies to iteratively refine motion representations and integrates a specialised cross self-attention mechanism to enhance multi-scale spatio-temporal feature extraction. When evaluated on the K400, UCF101, and HMDB51 datasets, RETIT achieved top-1 accuracies of 72.1%, 97.6%, and 75.8%, respectively, surpassing existing state-of-the-art benchmarks. These results highlight RETIT's ability to effectively leverage spatial and temporal modalities, advancing the state of action recognition in compressed video formats.

Keywords: compressed video analysis; action recognition; transformer architecture; spatio-temporal features; recursive frame interaction.

DOI: 10.1504/IJBIC.2026.153418

International Journal of Bio-Inspired Computation, 2026 Vol.27 No.3, pp.169 - 178

Received: 10 Oct 2024
Accepted: 25 Jan 2025
Published online: 07 May 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article