Open robertjyh opened 3 hours ago
I have not tested the translation effect of this dataset, my research direction is only in the direction of sign language recognition, and actually the recognition effect is not good, the WER of the test set is only 41.9%. The spoken translation is given to describe clearly how this dataset is built in the process. I know a way to improve the recognition performance dramatically, you can improve the model performance dramatically by pre-training the model in CSL-Daily first, and then train it in CE-CSL dataset, the recognition WER of the test set can reach about 32%, I don't know if this method is effective for you. Thanks!
------------------ 原始邮件 ------------------ 发件人: "woshisad159/TFNet" @.>; 发送时间: 2024年11月2日(星期六) 晚上7:33 @.>; @.***>; 主题: [woshisad159/TFNet] About translation effect (Issue #1)
Your dataset makes sign language tasks closer to real application scenarios. Thank you for your contribution! In real environments, in addition to recognition tasks, we also need to face translation tasks. I used your dataset to train a translation model, but the effect was very poor. This may be because the dataset itself is difficult or my method is not reasonable enough. Have you tested the translation effect on this dataset? Thanks!
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>
Thank you for your supplement! I will try to use CSL-Daily to improve my training effect.
Your dataset makes sign language tasks closer to real application scenarios. Thank you for your contribution! In real environments, in addition to recognition tasks, we also need to face translation tasks. I used your dataset to train a translation model, but the effect was very poor. This may be because the dataset itself is difficult or my method is not reasonable enough. Have you tested the translation effect on this dataset? Thanks!