Denoising-Diffusion Alignment for Continuous Sign Language Recognition

As a key to social good, continuous sign language recognition (CSLR) aims to promote active and accessible communication for the hearing impaired. Current CSLR research adopts a cross-modality alignment scheme to learn the mapping relationship between "video clip-textual gloss". However, this local alignment method, especially with weak data annotation, ignores the contextual information of modalities and directly reduces the generalization of visual features. To this end, we propose a novel Denoising-Diffusion global Alignment scheme (DDA), which focuses on modeling the mapping of the "entire video-gloss sequence". DDA consists of a partial noising process strategy and a denoising-diffusion autoencoder. The former is used to achieve efficient guidance of the text modality to the visual modality; the latter learns the global alignment information of the two modalities in a denoising manner. Our DDA confirms the feasibility of diffusion models for visual representation learning in CSLR. Experiments on three public benchmarks demonstrate that our method achieves state-of-the-art performances. Furthermore, the proposed method can be a plug-and-play optimization to generalize other CSLR methods.