Uni-Sign: Toward Unified Sign Language Understanding at Scale

Authors: Zecheng Li, Wengang Zhou, Weichao Zhao, Kepeng Wu, Hezhen Hu, Houqiang Li

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments across multiple SLU benchmarks demonstrate that Uni-Sign achieves state-of-the-art performance across multiple downstream SLU tasks.
Researcher Affiliation Academia 1 Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China 2 Institute of Artificial Intelligence, Hefei Comprehensive National Science Center 3 University of Texas at Austin EMAIL EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Pseudocode of the score-aware sampling strategy in a Py Torch-like style.
Open Source Code Yes Dataset and code are available at github.com/ZechengLi19/Uni-Sign.
Open Datasets Yes We introduce CSL-News, a large-scale Chinese Sign Language (CSL) dataset containing 1,985 hours of video paired with textual annotations, which enables effective large-scale pre-training. Dataset and code are available at github.com/ZechengLi19/Uni-Sign.
Dataset Splits Yes For ISLR, we adopt WLASL (Li et al., 2020a) and MSASL (Joze & Koller, 2019) datasets for evaluation. For CSLR, we utilize CSL-Daily (Zhou et al., 2021). SLT task is conducted on the CSL-Daily, How2Sign (Duarte et al., 2021), and Open ASL (Shi et al., 2022) datasets. (Tables 4, 5, and 6 show results for 'Dev' and 'Test' splits for these datasets).
Hardware Specification No It was also supported by the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC.
Software Dependencies No We implement Uni-Sign using Py Torch (Paszke et al., 2019), employing m T5Base (Xue et al., 2021) as our pre-trained language model.
Experiment Setup Yes The detailed training recipe is presented in Table 2. Config Stage 1 Stage 2 Stage 3 optimizer Adam W base learning rate 3e-4 weight decay 1e-4 optimizer momentum β1, β2=0.9, 0.999 learning rate schedule cosine decay training epochs 20 5 20 batch size 16 4 8 gradient accumulation 8 8 1