Hierarchical LSTM for Sign Language Translation

Authors: Dan Guo, Wengang Zhou, Houqiang Li, Meng Wang

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiment Experiment Setup Dataset: Our dataset is a collection of videos covering 100 daily sentences in Chinese sign language (CSL)1. Each sentence is played by 50 signers. It contains 50*100=5000 videos. The vocabulary size is 179. Each sentence contains 4 8 (average 5) sign words (phases). To validate our method, we split the dataset by two strategies shown in Table 2.
Researcher Affiliation Academia Dan Guo,1 Wengang Zhou,2 Houqiang Li,2 Meng Wang1 1School of Computer and Information, Hefei University of Technology, Hefei, 230009, China 2EEIS Department, University of Science and Technology of China, Hefei, 230027, China
Pseudocode No The paper describes methods in text and equations but does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or link indicating that the authors' own source code for the proposed HLSTM model is publicly available.
Open Datasets Yes Our dataset: Our dataset is a collection of videos covering 100 daily sentences in Chinese sign language (CSL)1. Each sentence is played by 50 signers. It contains 50*100=5000 videos... 1http://mccipc.ustc.edu.cn/mediawiki/index.php/SLR Dataset
Dataset Splits Yes To validate our method, we split the dataset by two strategies shown in Table 2. (a) Split I signer independent test: It splits video samples of 40 signers as training set and that of the remaining 10 signers as test set. The sentences of training and test sets are the same, but the signers are different. (b) Split II unseen sentences test: This strategy elaborately selects up to 6 sentences as test set, and the left 94 sentences as training set.
Hardware Specification No The paper does not provide any specific hardware details (e.g., GPU/CPU models, memory specifications) used for running the experiments.
Software Dependencies No The paper mentions software like 'C3D model', 'LSTM&CTC (Warp-ctc)', 'VGG', and 'LSTM' but does not specify any version numbers for these or other software dependencies.
Experiment Setup Yes Experiment on different LSTM hidden state numbers: To test the precision with different LSTM unit settings, we set nhid to 256, 512 and 1000, respectively. As shown in Table 5, when the nhid is set larger, the precision is raised. What s more, by our observation, when nhid is small, experimental results are instable under multiple random tests. However, when nhid = 1000, the results are stable. Thus we select the nhid = 1000 as our LSTM parameter setting. Experiment on different pooling strategies: From Table 6, the result accords with properties of these pooling strategies. ...Thus we separately set mean and max pooling for Split I and II.