This is the authors' abstract. We don't add key points for this paper.
In this task, we propose a video frame sequence processing module for the isolated word lip-reading task. Although 3D convolution is widely used for processing three-dimensional input data and can capture both temporal and spatial information, it does not effectively address the issue of capturing inter-frame correlations. To solve this problem, we introduce a multi-head recurrent attention mechanism, which effectively integrates inter-frame features from all video frames and further enhances attention to each frame by exploring inter-frame relationships. Additionally, we used the squeeze-and-excitation block, a lightweight attention mechanism, to further improve the model's classification capability. Our method achieved an accuracy of 88.91% on the Lip Reading in the Wild (LRW) dataset and 56.99% on the LRW-1000 dataset, comparable to or even surpassing existing state-of-the-art results.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Hu et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: